Stocks were higher as traders weighed disappointing earnings against new data that gave hope of easier Federal Reserve policy later this year.

Dow closes more than 300 points higher, securing 7th winning day: Live updates

Stocks have lost their bounce after a growing chorus of Federal Reserve officials made it clear that interest rates will need to stay higher for longer.

Stock market today: Stocks rise as Dow tries to extend 6-day win streak

The stock market climbed to its highest level since early April, extending a rebound that’s been fueled by speculation the Federal Reserve will be able to cut interest rates this year.

Stock Market Today: Dow, S&P Live Updates for May 9 - Bloomberg

Results from Disney took center stage as the media giant posted the first profit for a key part of its streaming business.

Stock market today: S&P 500 rises for 4th day, Disney sinks 9% after earnings

Mammoth will suck in air using giant fans, separate the carbon and transport it underground where it will transform into stone

‘World’s largest’ plant to suck carbon out of the air and turn it into stone opens in Iceland | CNN

Cabin crew taking sudden sick leave have led to flight disruptions at Tata Group budget carrier Air India Express since Tuesday night, the company said in a statement on Wednesday.

Air India Express says flights disrupted after sudden sick leave by cabin crew | Reuters

After months of speculation, Apple&#x27;s new iPads have finally dropped. But one major question remains.

The New iPad Air Is Apple’s Most Important Release In Years

CNBC Pro screened FactSet for stocks from four exchange-traded funds. Here are eight that turned up on the screen.

These global tech stocks have beaten the S&P 500 this year — and analysts give them over 20% upside

The Regional Labour Commissioner had written to Air India express and, in no uncertain terms, told them that the employees&#x27; grievances were genuine and that the HR department had made an attempt to mislead the conciliation officer.

Air India Express Crisis: Labour Commissioner Slams Tatas, Accuses Air India Express Of Mismanagement

Aair

aair

Authors:
(1) Qian Yang, Zhejiang University, Equal contribution. This work was conducted during Qian Yang’s internship at Alibaba Group;
(2) Jin Xu, Alibaba Group, Equal contribution;
(3) Wenrui Liu, Zhejiang University;
(4) Yunfei Chu, Alibaba Group;
(5) Xiaohuan Zhou, Alibaba Group;
(6) Yichong Leng, Alibaba Group;
(7) Yuanjun Lv, Alibaba Group;
(8) Zhou Zhao, Alibaba Group and Corresponding to Zhou Zhao (zhaozhou@zju.edu.cn);
(9) Yichong Leng, Zhejiang University
(10) Chang Zhou, Alibaba Group and Corresponding to Chang Zhou (ericzhou.zc@alibaba-inc.com);
(11) Jingren Zhou, Alibaba Group. Table of Links Abstract and 1. Introduction 2 Related Work 3 AIR-Bench and 3.1 Overview 3.2 Foundation Benchmark 3.3 Chat Benchmark 3.4 Evaluation Strategy 4 Experiments 4.1 Models 4.2 Main Results 4.3 Human Evaluation and 4.4 Ablation Study of Positional Bias 5 Conclusion and References A Detailed Results of Foundation Benchmark 3.3 Chat Benchmark Data Source and Audio Mixing Strategy. As shown in Table 2, we have collected more than 2k data samples spanning various audio types including speech, sound, music, and mixed audio. The purpose of introducing mixed audio is to augment the complexity of the audio signals and make it closer to audio from real-world audio scenarios. To achieve this, we propose a novel mixing strategy involving loudness control and temporal dislocation, as illustrated in Fig. 2. Specifically, we can adjust the relative loudness and temporal relationship between two audio clips for mixing. Then, we can create a complex audio signal that combines their meta-information, such as speech transcription accompanied by a background music caption. Furthermore, the meta-information also includes labels indicating which audio clip is louder and which is ahead in the temporal sequence. Open-ended Query and Reference. To prompt GPT-4 to generate open-ended question-answer pairs for audio, we should interpret the rich information in each audio with texts. We collect all of meta-information such as gender, age, emotion, transcription, language for speech, caption for natural sound, and instrument, caption for music from the original dataset. Rather than relying on pretrained models to extract this meta-information for each audio clip, we adopt the ground truth metainformation to avoid potential errors. After gathering meta-information about the audio, we manually construct prompts. These prompts are carefully designed to ensure a comprehensive coverage of chat interactions, taking into consideration the diverse range of audio signals involved. We design the prompts to facilitate the generation of questions related to the perception and reasoning for different types of audio. For the natural sound, the prompts are further tailored to generate questions that involve determining appropriate responses to sound events within a specific scenario. For the music category, prompts are devised to elicit creative writing and story generation questions based on music composition. To ensure the quality of the generated results, these prompts are designed in a manner that enables GPT-4 to automatically filter out responses that are not directly related to audio. Additionally, we manually reviewed all the question-answer pairs to ensure the quality of the questions and the reliability of the answers. The generated answers from GPT-4 are considered as references. This paper is available on arxiv under CC BY 4.0 DEED license. Authors: (1) Qian Yang, Zhejiang University, Equal contribution. This work was conducted during Qian Yang’s internship at Alibaba Group; (2) Jin Xu, Alibaba Group, Equal contribution; (3) Wenrui Liu, Zhejiang University; (4) Yunfei Chu, Alibaba Group; (5) Xiaohuan Zhou, Alibaba Group; (6) Yichong Leng, Alibaba Group; (7) Yuanjun Lv, Alibaba Group; (8) Zhou Zhao, Alibaba Group and Corresponding to Zhou Zhao (zhaozhou@zju.edu.cn); (9) Yichong Leng, Zhejiang University (10) Chang Zhou, Alibaba Group and Corresponding to Chang Zhou (ericzhou.zc@alibaba-inc.com); (11) Jingren Zhou, Alibaba Group. Authors: Authors: (1) Qian Yang, Zhejiang University, Equal contribution. This work was conducted during Qian Yang’s internship at Alibaba Group; (2) Jin Xu, Alibaba Group, Equal contribution; (3) Wenrui Liu, Zhejiang University; (4) Yunfei Chu, Alibaba Group; (5) Xiaohuan Zhou, Alibaba Group; (6) Yichong Leng, Alibaba Group; (7) Yuanjun Lv, Alibaba Group; (8) Zhou Zhao, Alibaba Group and Corresponding to Zhou Zhao (zhaozhou@zju.edu.cn); (9) Yichong Leng, Zhejiang University (10) Chang Zhou, Alibaba Group and Corresponding to Chang Zhou (ericzhou.zc@alibaba-inc.com); (11) Jingren Zhou, Alibaba Group. Table of Links Abstract and 1. Introduction Abstract and 1. Introduction 2 Related Work 2 Related Work 3 AIR-Bench and 3.1 Overview 3 AIR-Bench and 3.1 Overview 3.2 Foundation Benchmark 3.2 Foundation Benchmark 3.3 Chat Benchmark 3.3 Chat Benchmark 3.4 Evaluation Strategy 3.4 Evaluation Strategy 4 Experiments 4.1 Models 4.1 Models 4.2 Main Results 4.2 Main Results 4.3 Human Evaluation and 4.4 Ablation Study of Positional Bias 4.3 Human Evaluation and 4.4 Ablation Study of Positional Bias 5 Conclusion and References 5 Conclusion and References A Detailed Results of Foundation Benchmark A Detailed Results of Foundation Benchmark 3.3 Chat Benchmark Data Source and Audio Mixing Strategy. As shown in Table 2, we have collected more than 2k data samples spanning various audio types including speech, sound, music, and mixed audio. The purpose of introducing mixed audio is to augment the complexity of the audio signals and make it closer to audio from real-world audio scenarios. To achieve this, we propose a novel mixing strategy involving loudness control and temporal dislocation, as illustrated in Fig. 2. Specifically, we can adjust the relative loudness and temporal relationship between two audio clips for mixing. Then, we can create a complex audio signal that combines their meta-information, such as speech transcription accompanied by a background music caption. Furthermore, the meta-information also includes labels indicating which audio clip is louder and which is ahead in the temporal sequence. Data Source and Audio Mixing Strategy. Open-ended Query and Reference. To prompt GPT-4 to generate open-ended question-answer pairs for audio, we should interpret the rich information in each audio with texts. We collect all of meta-information such as gender, age, emotion, transcription, language for speech, caption for natural sound, and instrument, caption for music from the original dataset. Rather than relying on pretrained models to extract this meta-information for each audio clip, we adopt the ground truth metainformation to avoid potential errors. Open-ended Query and Reference. After gathering meta-information about the audio, we manually construct prompts. These prompts are carefully designed to ensure a comprehensive coverage of chat interactions, taking into consideration the diverse range of audio signals involved. We design the prompts to facilitate the generation of questions related to the perception and reasoning for different types of audio. For the natural sound, the prompts are further tailored to generate questions that involve determining appropriate responses to sound events within a specific scenario. For the music category, prompts are devised to elicit creative writing and story generation questions based on music composition. To ensure the quality of the generated results, these prompts are designed in a manner that enables GPT-4 to automatically filter out responses that are not directly related to audio. Additionally, we manually reviewed all the question-answer pairs to ensure the quality of the questions and the reliability of the answers. The generated answers from GPT-4 are considered as references. This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv

AIR-Bench’s Chat Benchmark: Open-Ended Audio QA with Mixed Audio Complexity

Authors:
(1) Qian Yang, Zhejiang University, Equal contribution. This work was conducted during Qian Yang’s internship at Alibaba Group;
(2) Jin Xu, Alibaba Group, Equal contribution;
(3) Wenrui Liu, Zhejiang University;
(4) Yunfei Chu, Alibaba Group;
(5) Xiaohuan Zhou, Alibaba Group;
(6) Yichong Leng, Alibaba Group;
(7) Yuanjun Lv, Alibaba Group;
(8) Zhou Zhao, Alibaba Group and Corresponding to Zhou Zhao (zhaozhou@zju.edu.cn);
(9) Yichong Leng, Zhejiang University
(10) Chang Zhou, Alibaba Group and Corresponding to Chang Zhou (ericzhou.zc@alibaba-inc.com);
(11) Jingren Zhou, Alibaba Group. Table of Links Abstract and 1. Introduction 2 Related Work 3 AIR-Bench and 3.1 Overview 3.2 Foundation Benchmark 3.3 Chat Benchmark 3.4 Evaluation Strategy 4 Experiments 4.1 Models 4.2 Main Results 4.3 Human Evaluation and 4.4 Ablation Study of Positional Bias 5 Conclusion and References A Detailed Results of Foundation Benchmark 3 AIR-Bench There exist three unique characteristics that differentiate AIR-Bench from existing benchmarks for audio understanding: i) AIR-Bench is the first work to incorporate task evaluation from all types of audio in a hierarchical taxonomy; ii) AIR-Bench is the first generative evaluation benchmark that handles the free-form output of LALMs; iii) AIRBench adopts GPT-4-based automatic evaluation yielding trustworthy evaluation results with affordable cost. In Sec. 3.1, we present the hierarchical taxonomy of AIR-Bench and discuss the design philosophy behind it. In Sec. 3.2 and Sec. 3.3, we introduce how we collect the audio-central question-answer pairs for foundation and chat tasks. In Sec. 3.4, we present the evaluation framework of AIR-Bench. 3.1 Overview Chat interaction based on audio is a complex task that encompasses a variety of fundamental competencies. For instance, humans are able to respond to sound events due to their capacities for sound perception and common sense reasoning. Similarly, the ability to respond to others’ spoken words is predicated on foundational skills such as speech-totext recognition and emotion recognition. Based on the motivation, we propose the hierarchical benchmark AIR-Bench by dividing it into foundation and chat benchmarks. The fundamental one is designed to assess capabilities across individual subtasks, serving to diagnose weaknesses within the model, while the chat benchmark directly evaluates complicated audio-based open-ended questions. The data sample is denoted as (A, Q, R), where A denotes the audio, Q represents the query and R is the reference answer. • Foundation benchmark: The purpose of the benchmark is to evaluate the individual capabilities of foundational tasks. To reduce the task difficulties and enable the evaluation of various models, we utilize the single-choice question-answering format. Specifically, the query Q is formed by concatenating a question q and candidate choices C, denoted as Q = (q, C). We curate a collection of 19 audio tasks that span multiple audio types, such as speech, music, and sound. These tasks include tasks like emotion recognition, acoustic scene classification, and music QA. [1] • Chat benchmark: The benchmark encompasses any form of question and answer pairs that could arise from audio signals, with the aim of reflecting the model’s ability to genuinely follow user instructions to perform perceiving, reasoning and interacting within realworld applications. According to the type of audio, the benchmark is categorized into four dimensions: speech, sound, music, and mixed audio, where mixed audio refers to audio that is a mixture of multiple types of audio, such as human voice with background music. The overview of AIR-Bench is shown in Fig. 1. This paper is available on arxiv under CC BY 4.0 DEED license. [1] For transcription tasks such as ASR and S2TT, we incorporate them into the chat benchmark since they are not suitable for the single-choice task format. Authors: (1) Qian Yang, Zhejiang University, Equal contribution. This work was conducted during Qian Yang’s internship at Alibaba Group; (2) Jin Xu, Alibaba Group, Equal contribution; (3) Wenrui Liu, Zhejiang University; (4) Yunfei Chu, Alibaba Group; (5) Xiaohuan Zhou, Alibaba Group; (6) Yichong Leng, Alibaba Group; (7) Yuanjun Lv, Alibaba Group; (8) Zhou Zhao, Alibaba Group and Corresponding to Zhou Zhao (zhaozhou@zju.edu.cn); (9) Yichong Leng, Zhejiang University (10) Chang Zhou, Alibaba Group and Corresponding to Chang Zhou (ericzhou.zc@alibaba-inc.com); (11) Jingren Zhou, Alibaba Group. Authors: Authors: (1) Qian Yang, Zhejiang University, Equal contribution. This work was conducted during Qian Yang’s internship at Alibaba Group; (2) Jin Xu, Alibaba Group, Equal contribution; (3) Wenrui Liu, Zhejiang University; (4) Yunfei Chu, Alibaba Group; (5) Xiaohuan Zhou, Alibaba Group; (6) Yichong Leng, Alibaba Group; (7) Yuanjun Lv, Alibaba Group; (8) Zhou Zhao, Alibaba Group and Corresponding to Zhou Zhao (zhaozhou@zju.edu.cn); (9) Yichong Leng, Zhejiang University (10) Chang Zhou, Alibaba Group and Corresponding to Chang Zhou (ericzhou.zc@alibaba-inc.com); (11) Jingren Zhou, Alibaba Group. Table of Links Abstract and 1. Introduction Abstract and 1. Introduction 2 Related Work 2 Related Work 3 AIR-Bench and 3.1 Overview 3 AIR-Bench and 3.1 Overview 3.2 Foundation Benchmark 3.2 Foundation Benchmark 3.3 Chat Benchmark 3.3 Chat Benchmark 3.4 Evaluation Strategy 3.4 Evaluation Strategy 4 Experiments 4.1 Models 4.1 Models 4.2 Main Results 4.2 Main Results 4.3 Human Evaluation and 4.4 Ablation Study of Positional Bias 4.3 Human Evaluation and 4.4 Ablation Study of Positional Bias 5 Conclusion and References 5 Conclusion and References A Detailed Results of Foundation Benchmark A Detailed Results of Foundation Benchmark 3 AIR-Bench There exist three unique characteristics that differentiate AIR-Bench from existing benchmarks for audio understanding: i) AIR-Bench is the first work to incorporate task evaluation from all types of audio in a hierarchical taxonomy; ii) AIR-Bench is the first generative evaluation benchmark that handles the free-form output of LALMs; iii) AIRBench adopts GPT-4-based automatic evaluation yielding trustworthy evaluation results with affordable cost. In Sec. 3.1, we present the hierarchical taxonomy of AIR-Bench and discuss the design philosophy behind it. In Sec. 3.2 and Sec. 3.3, we introduce how we collect the audio-central question-answer pairs for foundation and chat tasks. In Sec. 3.4, we present the evaluation framework of AIR-Bench. 3.1 Overview Chat interaction based on audio is a complex task that encompasses a variety of fundamental competencies. For instance, humans are able to respond to sound events due to their capacities for sound perception and common sense reasoning. Similarly, the ability to respond to others’ spoken words is predicated on foundational skills such as speech-totext recognition and emotion recognition. Based on the motivation, we propose the hierarchical benchmark AIR-Bench by dividing it into foundation and chat benchmarks. The fundamental one is designed to assess capabilities across individual subtasks, serving to diagnose weaknesses within the model, while the chat benchmark directly evaluates complicated audio-based open-ended questions. The data sample is denoted as (A, Q, R), where A denotes the audio, Q represents the query and R is the reference answer. reference answer. • Foundation benchmark: The purpose of the benchmark is to evaluate the individual capabilities of foundational tasks. To reduce the task difficulties and enable the evaluation of various models, we utilize the single-choice question-answering format. Specifically, the query Q is formed by concatenating a question q and candidate choices C, denoted as Q = (q, C). We curate a collection of 19 audio tasks that span multiple audio types, such as speech, music, and sound. These tasks include tasks like emotion recognition, acoustic scene classification, and music QA. [1] • Foundation benchmark: • Chat benchmark: The benchmark encompasses any form of question and answer pairs that could arise from audio signals, with the aim of reflecting the model’s ability to genuinely follow user instructions to perform perceiving, reasoning and interacting within realworld applications. According to the type of audio, the benchmark is categorized into four dimensions: speech, sound, music, and mixed audio, where mixed audio refers to audio that is a mixture of multiple types of audio, such as human voice with background music. • Chat benchmark: The overview of AIR-Bench is shown in Fig. 1. This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv [1] For transcription tasks such as ASR and S2TT, we incorporate them into the chat benchmark since they are not suitable for the single-choice task format.

What is AIR-Bench?

Authors:
(1) Qian Yang, Zhejiang University, Equal contribution. This work was conducted during Qian Yang’s internship at Alibaba Group;
(2) Jin Xu, Alibaba Group, Equal contribution;
(3) Wenrui Liu, Zhejiang University;
(4) Yunfei Chu, Alibaba Group;
(5) Xiaohuan Zhou, Alibaba Group;
(6) Yichong Leng, Alibaba Group;
(7) Yuanjun Lv, Alibaba Group;
(8) Zhou Zhao, Alibaba Group and Corresponding to Zhou Zhao (zhaozhou@zju.edu.cn);
(9) Yichong Leng, Zhejiang University
(10) Chang Zhou, Alibaba Group and Corresponding to Chang Zhou (ericzhou.zc@alibaba-inc.com);
(11) Jingren Zhou, Alibaba Group. Table of Links Abstract and 1. Introduction 2 Related Work 3 AIR-Bench and 3.1 Overview 3.2 Foundation Benchmark 3.3 Chat Benchmark 3.4 Evaluation Strategy 4 Experiments 4.1 Models 4.2 Main Results 4.3 Human Evaluation and 4.4 Ablation Study of Positional Bias 5 Conclusion and References A Detailed Results of Foundation Benchmark Abstract Recently, instruction-following audio-language models have received broad attention for human-audio interaction. However, the absence of benchmarks capable of evaluating audio-centric interaction capabilities has impeded advancements in this field. Previous models primarily focus on assessing different fundamental tasks, such as Automatic Speech Recognition (ASR), and lack an assessment of the open-ended generative capabilities centered around audio. Thus, it is challenging to track the progression in the Large Audio-Language Models (LALMs) domain and to provide guidance for future improvement. In this paper, we introduce AIR-Bench (Audio InstRuction Benchmark), the first benchmark designed to evaluate the ability of LALMs to understand various types of audio signals (including human speech, natural sounds and music), and furthermore, to interact with humans in textual format. AIR-Bench encompasses two dimensions: foundation and chat benchmarks. The former consists of 19 tasks with approximately 19k single-choice questions, intending to inspect the basic single-task ability of LALMs. The latter one contains 2k instances of open-ended question-and-answer data, directly assessing the comprehension of the model on complex audio and its capacity to follow instructions. Both benchmarks require the model to generate hypotheses directly. We design a unified framework that leverages advanced language models, such as GPT-4, to evaluate the scores of generated hypotheses given the meta-information of the audio. Experimental results demonstrate a high level of consistency between GPT-4 based evaluation and human evaluation. By revealing the limitations of existing LALMs through evaluation results, AIR-Bench can provide insights on the direction of future research. 1 Introduction Recent advancements in artificial general intelligence have been significantly driven by the emergence of large language models (LLMs) (Brown et al., 2020; OpenAI, 2022, 2023; Chowdhery et al., 2022; Anil et al., 2023; Touvron et al., 2023a,b; Bai et al., 2023a). These models exhibit remarkable abilities in retaining knowledge, engaging in intricate reasoning, and solving problems following human intents. Motivated by the striking progress in large language models (LLMs), the domain of large audio-language models (LALMs) has undergone a revolutionary transformation. To perceive and comprehend rich audio signals and further generate textual responses following human instructions, many works have been proposed, such as SALMONN (Tang et al., 2023a), BLSP (Wang et al., 2023a), Speech-LLaMA (Wu et al., 2023a), and Qwen-Audio (Chu et al., 2023), showcasing promising capabilities for audio-central dialogues. However, previous LALMs (Tang et al., 2023a; Wang et al., 2023a; Wu et al., 2023a; Chu et al., 2023; Huang et al., 2023b; Shen et al., 2023; Gong et al., 2023; Wang et al., 2023b) have predominantly concentrated on evaluation in specific fundamental tasks. The absence of a standardized benchmark for assessing the generative instructionfollowing abilities of these models has resulted in a reliance on showcasing examples or releasing the chat models for public experimentation to demonstrate their conversational skills. This approach poses significant challenges for conducting fair and objective comparisons across different research endeavors. Moreover, it tends to obscure the models’ existing limitations, impeding the ability to monitor advancements within the domain of LALMs. For evaluation in audio domains, the majority of research efforts have concentrated on the creation of benchmarks tailored to individual tasks such as LibriSpeech (Panayotov et al., 2015) and Common Voice benchmark (Ardila et al., 2019) for ASR. Beyond task-specific ones, benchmarks like SUPERB (Yang et al., 2021a) and HEAR (Turian et al., 2021) have been designed to test the versatility of self-supervised learning models in a wide variety of tasks. Regarding the assessment of LALMs’ ability to follow instructions, to the best of our knowledge, Dynamic-SUPERB (Huang et al., 2023a) is the only benchmark devoted to this aspect. Nevertheless, Dynamic-SUPERB only focuses on human speech processing, and does not extend to the assessment of models’ capabilities in producing open-ended generations such as dialogues. In this paper, we present AIR-Bench (Audio InstRuction Benchmark), a novel benchmark designed to evaluate the ability of LALMs to comprehend various audio signals and to interact following instructions. AIR-Bench is characterized by three primary features: 1) Comprehensive audio signals coverage. AIR-Bench offers comprehensive coverage of audio signals, including human speech, natural sounds, and music, ensuring a comprehensive evaluation of LALMs’ capabilities. 2) Hierarchical Benchmark Structure. The benchmark consists of foundation and chat benchmarks. The foundation benchmark comprises 19 distinct audio tasks with over 19,000 single-choice questions, with each question focusing only on a specific foundational ability. GPT-4 (OpenAI, 2023) extends the questions and candidate choices using dedicated designed prompts. The chat component consists of over 2,000 audio-prompted open-ended questions. To enhance the complexity of the audio and achieve a closer resemblance to the intricate audio encountered in real-life situations, we propose a novel audio mixing strategy that incorporates loudness control and temporal dislocation. Specifically, we adjust the loudness and introduce different temporal offsets during the mixing process of two audio clips. The resulting variations in relative loudness and temporal location are then recorded as additional meta-information, contributing to a more comprehensive textual representation of the audio. The quality of data is upheld through automated filtering by GPT-4, followed by manual verification. 3) Unified, objective and reproducible evaluation framework. Models are required to generate hypothesis sequences directly across both benchmarks to align more accurately with practical scenarios. Then, we employ GPT-4 to generate reference answers given meta-information through carefully constructed prompts. Given references and hypotheses, following Liu et al. (2023b); Bai et al. (2023b), we use GPT-4 (OpenAI, 2023) to judge whether the choice is correct for the foundation benchmark or score hypotheses for the chat benchmark. We further perform a second scoring by swapping their positions to eliminate the position bias. Based on comprehensive experiments on 9 LALMs, we observe that existing LALMs either have limited audio understanding or instruction-following capabilities, leaving significant room for improvement in this field. Our contribution is summarized below: • AIR-Bench is the first generative evaluation benchmark for large audio-language models, encompassing a wide array of audio such as speech, natural sounds and music. AIR-Bench is a large and hierarchical benchmark, consisting of the foundation benchmark with 19 audio tasks and over 19k single-choice questions, alongside a chat benchmark with over 2k meticulously curated open-ended audio questions for comprehensive evaluation. • We propose a novel audio mixing strategy with loudness control and temporal dislocation to enhance the complexity of the audio. • A unified, objective, and reproducible evaluation framework has been developed to assess the quality of generative hypotheses. • We conducted a thorough evaluation of 9 models for the purpose of benchmarking. The evaluation code, datasets, and an open leaderboard will be made publicly available soon. This paper is available on arxiv under CC BY 4.0 DEED license. Authors: (1) Qian Yang, Zhejiang University, Equal contribution. This work was conducted during Qian Yang’s internship at Alibaba Group; (2) Jin Xu, Alibaba Group, Equal contribution; (3) Wenrui Liu, Zhejiang University; (4) Yunfei Chu, Alibaba Group; (5) Xiaohuan Zhou, Alibaba Group; (6) Yichong Leng, Alibaba Group; (7) Yuanjun Lv, Alibaba Group; (8) Zhou Zhao, Alibaba Group and Corresponding to Zhou Zhao (zhaozhou@zju.edu.cn); (9) Yichong Leng, Zhejiang University (10) Chang Zhou, Alibaba Group and Corresponding to Chang Zhou (ericzhou.zc@alibaba-inc.com); (11) Jingren Zhou, Alibaba Group. Authors: Authors: (1) Qian Yang, Zhejiang University, Equal contribution. This work was conducted during Qian Yang’s internship at Alibaba Group; (2) Jin Xu, Alibaba Group, Equal contribution; (3) Wenrui Liu, Zhejiang University; (4) Yunfei Chu, Alibaba Group; (5) Xiaohuan Zhou, Alibaba Group; (6) Yichong Leng, Alibaba Group; (7) Yuanjun Lv, Alibaba Group; (8) Zhou Zhao, Alibaba Group and Corresponding to Zhou Zhao (zhaozhou@zju.edu.cn); (9) Yichong Leng, Zhejiang University (10) Chang Zhou, Alibaba Group and Corresponding to Chang Zhou (ericzhou.zc@alibaba-inc.com); (11) Jingren Zhou, Alibaba Group. Table of Links Abstract and 1. Introduction Abstract and 1. Introduction 2 Related Work 2 Related Work 3 AIR-Bench and 3.1 Overview 3 AIR-Bench and 3.1 Overview 3.2 Foundation Benchmark 3.2 Foundation Benchmark 3.3 Chat Benchmark 3.3 Chat Benchmark 3.4 Evaluation Strategy 3.4 Evaluation Strategy 4 Experiments 4.1 Models 4.1 Models 4.2 Main Results 4.2 Main Results 4.3 Human Evaluation and 4.4 Ablation Study of Positional Bias 4.3 Human Evaluation and 4.4 Ablation Study of Positional Bias 5 Conclusion and References 5 Conclusion and References A Detailed Results of Foundation Benchmark A Detailed Results of Foundation Benchmark Abstract Recently, instruction-following audio-language models have received broad attention for human-audio interaction. However, the absence of benchmarks capable of evaluating audio-centric interaction capabilities has impeded advancements in this field. Previous models primarily focus on assessing different fundamental tasks, such as Automatic Speech Recognition (ASR), and lack an assessment of the open-ended generative capabilities centered around audio. Thus, it is challenging to track the progression in the Large Audio-Language Models (LALMs) domain and to provide guidance for future improvement. In this paper, we introduce AIR-Bench (Audio InstRuction Benchmark), the first benchmark designed to evaluate the ability of LALMs to understand various types of audio signals (including human speech, natural sounds and music), and furthermore, to interact with humans in textual format. AIR-Bench encompasses two dimensions: foundation and chat benchmarks. The former consists of 19 tasks with approximately 19k single-choice questions, intending to inspect the basic single-task ability of LALMs. The latter one contains 2k instances of open-ended question-and-answer data, directly assessing the comprehension of the model on complex audio and its capacity to follow instructions. Both benchmarks require the model to generate hypotheses directly. We design a unified framework that leverages advanced language models, such as GPT-4, to evaluate the scores of generated hypotheses given the meta-information of the audio. Experimental results demonstrate a high level of consistency between GPT-4 based evaluation and human evaluation. By revealing the limitations of existing LALMs through evaluation results, AIR-Bench can provide insights on the direction of future research. 1 Introduction Recent advancements in artificial general intelligence have been significantly driven by the emergence of large language models (LLMs) (Brown et al., 2020; OpenAI, 2022, 2023; Chowdhery et al., 2022; Anil et al., 2023; Touvron et al., 2023a,b; Bai et al., 2023a). These models exhibit remarkable abilities in retaining knowledge, engaging in intricate reasoning, and solving problems following human intents. Motivated by the striking progress in large language models (LLMs), the domain of large audio-language models (LALMs) has undergone a revolutionary transformation. To perceive and comprehend rich audio signals and further generate textual responses following human instructions, many works have been proposed, such as SALMONN (Tang et al., 2023a), BLSP (Wang et al., 2023a), Speech-LLaMA (Wu et al., 2023a), and Qwen-Audio (Chu et al., 2023), showcasing promising capabilities for audio-central dialogues. However, previous LALMs (Tang et al., 2023a; Wang et al., 2023a; Wu et al., 2023a; Chu et al., 2023; Huang et al., 2023b; Shen et al., 2023; Gong et al., 2023; Wang et al., 2023b) have predominantly concentrated on evaluation in specific fundamental tasks. The absence of a standardized benchmark for assessing the generative instructionfollowing abilities of these models has resulted in a reliance on showcasing examples or releasing the chat models for public experimentation to demonstrate their conversational skills. This approach poses significant challenges for conducting fair and objective comparisons across different research endeavors. Moreover, it tends to obscure the models’ existing limitations, impeding the ability to monitor advancements within the domain of LALMs. For evaluation in audio domains, the majority of research efforts have concentrated on the creation of benchmarks tailored to individual tasks such as LibriSpeech (Panayotov et al., 2015) and Common Voice benchmark (Ardila et al., 2019) for ASR. Beyond task-specific ones, benchmarks like SUPERB (Yang et al., 2021a) and HEAR (Turian et al., 2021) have been designed to test the versatility of self-supervised learning models in a wide variety of tasks. Regarding the assessment of LALMs’ ability to follow instructions, to the best of our knowledge, Dynamic-SUPERB (Huang et al., 2023a) is the only benchmark devoted to this aspect. Nevertheless, Dynamic-SUPERB only focuses on human speech processing, and does not extend to the assessment of models’ capabilities in producing open-ended generations such as dialogues. In this paper, we present AIR-Bench (Audio InstRuction Benchmark), a novel benchmark designed to evaluate the ability of LALMs to comprehend various audio signals and to interact following instructions. AIR-Bench is characterized by three primary features: 1) Comprehensive audio signals coverage. AIR-Bench offers comprehensive coverage of audio signals, including human speech, natural sounds, and music, ensuring a comprehensive evaluation of LALMs’ capabilities. 2) Hierarchical Benchmark Structure. The benchmark consists of foundation and chat benchmarks. The foundation benchmark comprises 19 distinct audio tasks with over 19,000 single-choice questions, with each question focusing only on a specific foundational ability. GPT-4 (OpenAI, 2023) extends the questions and candidate choices using dedicated designed prompts. The chat component consists of over 2,000 audio-prompted open-ended questions. To enhance the complexity of the audio and achieve a closer resemblance to the intricate audio encountered in real-life situations, we propose a novel audio mixing strategy that incorporates loudness control and temporal dislocation. Specifically, we adjust the loudness and introduce different temporal offsets during the mixing process of two audio clips. The resulting variations in relative loudness and temporal location are then recorded as additional meta-information, contributing to a more comprehensive textual representation of the audio. The quality of data is upheld through automated filtering by GPT-4, followed by manual verification. 3) Unified, objective and reproducible evaluation framework. Models are required to generate hypothesis sequences directly across both benchmarks to align more accurately with practical scenarios. Then, we employ GPT-4 to generate reference answers given meta-information through carefully constructed prompts. Given references and hypotheses, following Liu et al. (2023b); Bai et al. (2023b), we use GPT-4 (OpenAI, 2023) to judge whether the choice is correct for the foundation benchmark or score hypotheses for the chat benchmark. We further perform a second scoring by swapping their positions to eliminate the position bias. Based on comprehensive experiments on 9 LALMs, we observe that existing LALMs either have limited audio understanding or instruction-following capabilities, leaving significant room for improvement in this field. Our contribution is summarized below: • AIR-Bench is the first generative evaluation benchmark for large audio-language models, encompassing a wide array of audio such as speech, natural sounds and music. AIR-Bench is a large and hierarchical benchmark, consisting of the foundation benchmark with 19 audio tasks and over 19k single-choice questions, alongside a chat benchmark with over 2k meticulously curated open-ended audio questions for comprehensive evaluation. • We propose a novel audio mixing strategy with loudness control and temporal dislocation to enhance the complexity of the audio. • A unified, objective, and reproducible evaluation framework has been developed to assess the quality of generative hypotheses. • We conducted a thorough evaluation of 9 models for the purpose of benchmarking. The evaluation code, datasets, and an open leaderboard will be made publicly available soon. This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv

Benchmarking Large Audio-Language Models via Generative Comprehension

Generativ Anlama vasitəsilə Böyük Səs-Dil Modellərini müqayisə etmək

Evaluación comparativa de grandes modelos de audio y lenguaje mediante comprensión generativa

Benchmarking Audio-Hizkuntza Eredu Handiak ulermen sortzailearen bidez

生成的理解による大規模音声言語モデルのベンチマーク

Генеративті түсіну арқылы үлкен дыбыстық-тілдік үлгілерді салыстыру

Генеративдик түшүнүү аркылуу чоң аудио-тил моделдерин салыштыруу

Famaritana ny maodely amin'ny fiteny audio lehibe amin'ny alàlan'ny fahatakarana miteraka

Porównywanie dużych modeli języka audio za pomocą rozumienia generatywnego

بنچمارک کول د لوی آډیو - ژبې ماډلونه د تولیدي پوهاوي له لارې

Бенцхмаркинг великих аудио-језичких модела путем генеративног разумевања

Generative düşünmek arkaly uly ses-dil modellerini deňeşdirmek

Ukulinganisa Amamodeli Amakhulu Olimi Lomsindo Nge-Generative Comprehension

The Curse of Pygmalion - What Would An Ideal AR/VR Device Look Like?

পিগম্যালিয়নের অভিশাপ - একটি আদর্শ এআর/ভিআর ডিভাইস দেখতে কেমন হবে?

Der Fluch von Pygmalion – Wie würde ein ideales AR/VR-Gerät aussehen?

La maldición de Pigmalión: ¿cómo sería un dispositivo AR/VR ideal?

La malédiction de Pygmalion – À quoi ressemblerait un appareil AR/VR idéal ?

पिग्मेलियन का अभिशाप - एक आदर्श AR/VR डिवाइस कैसा दिखेगा?

ピグマリオンの呪い - 理想的な AR/VR デバイスとはどのようなものでしょうか?

피그말리온의 저주 - 이상적인 AR/VR 장치는 어떤 모습일까요?

A Maldição de Pigmalião – Como seria um dispositivo AR/VR ideal?

Проклятие Пигмалиона: как должно выглядеть идеальное AR/VR-устройство?

Pygmalion'un Laneti - İdeal Bir AR/VR Cihazı Nasıl Görünür?

Lời nguyền của Pygmalion - Thiết bị AR/VR lý tưởng sẽ trông như thế nào?

皮格马利翁的诅咒——理想的 AR/VR 设备是什么样的？

As someone with a keen interest in web3, I witnessed for the first time how what started as a low-key trend on the internet spilled over into a major sensation that has now come to the knowledge of millions of people in Nigeria. On the streets of Lagos and several other places in Nigeria, hordes of people fervently tapping on their phones’ screens are now a common sight. As of this writing, the word on the street is that there is a lot of free money to be made from participating in a tapping contest sponsored by TapSwap, a new cryptocurrency project on the Solana blockchain. Most of the netizens that have taken an interest in TapSwap might have probably read stories about the payouts some lucky Nigerians received from mining Notcoin–a recently launched cryptocurrency–before joining the contest. There is also the trending story of a Nigerian who purchased a 300k NGN air-conditioner from his Notcoin “profits.’ The general consensus is that the payouts are real, and it is left for everyone to decide whether to jump on the bandwagon or stay out of it. Hardly had the news of Notcoin payouts faded when TapSwap came on board offering rewards to participants tapping on a gold coin that has the letter T inscribed on it. All you had to do was get a referral from a participant, download a Telegram app, create an account, and start tapping on the gold coin that appeared on your screen. For many Nigerians, this was the easy part. But as participants were required to download a phantom wallet, there was confusion in the air as to what it was and how to generate a Solana address from it. It turned out that the vast majority of the Nigerian TapSwappers had not heard of a phantom wallet and coming across it for the first time was definitely going to inspire intense curiosity and lead many down the rabbit hole. For some, this might be a life-changing opportunity to take a plunge into web3 and cryptocurrency trading, starting from scratch. And for others, it might just appear to be another balderdash or even a scam. With a few days remaining before the launch of TapSwap, millions of Nigerians are still tapping on their screens and anticipating making money from the airdrop. Local media have been monitoring the situation and the Nigerian financial regulators are waiting to see the outcome of the exercise. It has yet to be concluded whether this is another MMM, a popular pyramid scheme to which many Nigerians lost millions of invested funds. As the tapping frenzy is approaching its climax, it is crucial to highlight its effect on the Nigerian populace, especially as there is an ongoing clampdown on Binance and other cryptocurrency exchanges in the country. It is worth noting that the Nigerian financial regulator, SEC, has recently increased the amount required to register a cryptocurrency company from 30 million Naira to 150 million Naira, with other stringent laid down rules in an effort to police the industry. Many willing investors might view this new policy as anti-crypto, and it may compel them to opt for a liberal and less expensive destination. But will the TapSwap Frenzy that is currently putting Nigeria on the crypto map make any difference? Crypto Awareness Since the tapping sensation began in Nigeria, I have taken a keen interest in its trajectory, and like millions of other people, I am anxiously waiting for the outcome. I have heard rumors that the team behind the project is anonymous and that they would fall short of their promises. But my training as a lawyer requires me to gather my facts and treat them nicely before drawing a conclusion. Nevertheless, it is not rocket science to predict that the successful outcome of the ongoing exercise will be a major win for the Nigerian crypto industry, restoring balance to the skewed reputation of cryptocurrency as well as leading thousands down the rabbit hole. Most especially, it will boost crypto awareness across the country and mark the starting point for mass participation in the crypto industry. There are two key reasons one must consider to understand how a tapping sensation like TapSwap can lead to a major economic upturn in developing countries. The first reason is the digital revolution. As the promises of a digital and more connected world continue to hold true, Africa is no longer willing to be left behind. Crypto adoption is on the rise and one of Nigeria, Kenya, and South Africa are bound to lead this revolution. Another factor to consider is the poverty of most developing countries. Bitcoin’s promises of financial inclusion and liberation are still a dream many people across the world hold on dearly to. Today, it is not unusual to read expert predictions of Bitcoin’s future price action on the internet. This sentiment is so strong and generally shared that events of the last few months have led many to renew their conviction about Bitcoin surging to $1 million in the near future. However, the sentiment towards Bitcoin and cryptocurrency changing the status quo is not generally shared. There are billions of people across the world who are not part of the crypto industry, which is why the most crucial question in cryptocurrency is ‘Wen mass adoption?’ To understand the evolution of the crypto industry across the world, one must pay close attention to crucial events like the TapSwap frenzy in Nigeria. As someone with a keen interest in web3, I witnessed for the first time how what started as a low-key trend on the internet spilled over into a major sensation that has now come to the knowledge of millions of people in Nigeria. On the streets of Lagos and several other places in Nigeria, hordes of people fervently tapping on their phones’ screens are now a common sight. As of this writing, the word on the street is that there is a lot of free money to be made from participating in a tapping contest sponsored by TapSwap, a new cryptocurrency project on the Solana blockchain. Most of the netizens that have taken an interest in TapSwap might have probably read stories about the payouts some lucky Nigerians received from mining Notcoin–a recently launched cryptocurrency–before joining the contest. There is also the trending story of a Nigerian who purchased a 300k NGN air-conditioner from his Notcoin “profits.’ The general consensus is that the payouts are real, and it is left for everyone to decide whether to jump on the bandwagon or stay out of it. Hardly had the news of Notcoin payouts faded when TapSwap came on board offering rewards to participants tapping on a gold coin that has the letter T inscribed on it. All you had to do was get a referral from a participant, download a Telegram app, create an account, and start tapping on the gold coin that appeared on your screen. For many Nigerians, this was the easy part. But as participants were required to download a phantom wallet, there was confusion in the air as to what it was and how to generate a Solana address from it. It turned out that the vast majority of the Nigerian TapSwappers had not heard of a phantom wallet and coming across it for the first time was definitely going to inspire intense curiosity and lead many down the rabbit hole. For some, this might be a life-changing opportunity to take a plunge into web3 and cryptocurrency trading, starting from scratch. And for others, it might just appear to be another balderdash or even a scam. With a few days remaining before the launch of TapSwap, millions of Nigerians are still tapping on their screens and anticipating making money from the airdrop. Local media have been monitoring the situation and the Nigerian financial regulators are waiting to see the outcome of the exercise. It has yet to be concluded whether this is another MMM, a popular pyramid scheme to which many Nigerians lost millions of invested funds. As the tapping frenzy is approaching its climax, it is crucial to highlight its effect on the Nigerian populace, especially as there is an ongoing clampdown on Binance and other cryptocurrency exchanges in the country. It is worth noting that the Nigerian financial regulator, SEC, has recently increased the amount required to register a cryptocurrency company from 30 million Naira to 150 million Naira, with other stringent laid down rules in an effort to police the industry. Many willing investors might view this new policy as anti-crypto, and it may compel them to opt for a liberal and less expensive destination. But will the TapSwap Frenzy that is currently putting Nigeria on the crypto map make any difference? Crypto Awareness Since the tapping sensation began in Nigeria, I have taken a keen interest in its trajectory, and like millions of other people, I am anxiously waiting for the outcome. I have heard rumors that the team behind the project is anonymous and that they would fall short of their promises. But my training as a lawyer requires me to gather my facts and treat them nicely before drawing a conclusion. Nevertheless, it is not rocket science to predict that the successful outcome of the ongoing exercise will be a major win for the Nigerian crypto industry, restoring balance to the skewed reputation of cryptocurrency as well as leading thousands down the rabbit hole. Most especially, it will boost crypto awareness across the country and mark the starting point for mass participation in the crypto industry. There are two key reasons one must consider to understand how a tapping sensation like TapSwap can lead to a major economic upturn in developing countries. The first reason is the digital revolution. As the promises of a digital and more connected world continue to hold true, Africa is no longer willing to be left behind. Crypto adoption is on the rise and one of Nigeria, Kenya, and South Africa are bound to lead this revolution. Another factor to consider is the poverty of most developing countries. Bitcoin’s promises of financial inclusion and liberation are still a dream many people across the world hold on dearly to. Today, it is not unusual to read expert predictions of Bitcoin’s future price action on the internet. This sentiment is so strong and generally shared that events of the last few months have led many to renew their conviction about Bitcoin surging to $1 million in the near future. However, the sentiment towards Bitcoin and cryptocurrency changing the status quo is not generally shared. There are billions of people across the world who are not part of the crypto industry, which is why the most crucial question in cryptocurrency is ‘Wen mass adoption?’ To understand the evolution of the crypto industry across the world, one must pay close attention to crucial events like the TapSwap frenzy in Nigeria.

The TapSwap Frenzy: A Boost For Crypto Awareness In Nigeria?

The Five Pillars of Untouchable Excellence: The Secret Weapon Most Entrepreneurs Neglect

Transcendence Book 9: Far-skies

-𝙸𝚗 𝚝𝚑𝚎 𝚋𝚎𝚐𝚒𝚗𝚗𝚒𝚗𝚐, 𝚝𝚑𝚎𝚛𝚎 𝚠𝚊𝚜 𝚘𝚗𝚎. -𝙳𝚎𝚍𝚒𝚌𝚊𝚝𝚎𝚍 𝚝𝚘 𝚝𝚑𝚎 𝚌𝚑𝚒𝚕𝚍𝚛𝚎𝚗 𝚠𝚑𝚘 𝚗𝚎𝚟𝚎𝚛 𝚏𝚘𝚞𝚗𝚍 𝚑𝚘𝚖𝚎. Year 2048. The first words I ever heard were in a facility for kids. We were five, the golden children of humankind. From the age of 2, we could fire darts with precision, and it wasn’t weird. We never had anyone new join us; we were always alone, together. Father called us geniuses; he called us creatures beyond humanity. Father never hid the truth from us, and he treated us with utmost respect. We never saw Mother. Not even once. We only knew she existed because Father mentioned her from time to time. Father loved us, and he told us that we were his only hope. Father hid nothing from us; he believed in being straightforward. “ ” Father said one day, as he stared pointedly at the five of us, me being some sort of leader, according to Father. Father was a large man with tribal marks of jagged lines adorning his face, giving him the look of a warrior, but the lab coat he always wore conflicted with his weathered face, making him look like a cross between a villain and a doting father. The world will one day see all you can do. Until then, forgive me. We didn’t know what we had to forgive Father for; he was always there when we needed him, along with two others. Father called them his underlings, yet he was so friendly with them; the camaraderie was beautiful. We were treated well and didn’t have a cause to question our existence. Names were not given to us, and we lived with codes. Father said names were outdated and called us in codes. I was 0i. Father said codes were vital because we were different from the rest of the people in the books we read. Father made us feed on Socrates, Plato, and so many other ancient philosophers. His favorite was Friedrich Nietzsche, who spoke about the abyss. Father spoon-fed us knowledge, saying we needed to understand how the world worked. “ ” You’re only useful when you choose to be. The power you wield lies in your willingness to wield it. Man is a creature of willpower, and once willpower is taken away, nothing remains of man. A fate worse than death. Father’s favorite quote was a personal one. He never ceased to remind us of what would happen if our willpower was taken away. Father was the only god we knew. We didn’t exactly worship him; we loved him. He was all we knew, the light in our overwhelming darkness. A man worthy of the title, father. 0ii, one of the girls in our midst. She was sharp-eyed, her mouth rarely opening up to speak. With an oval face and a perfect nose, 0ii became fascinating. I once asked Father why she wasn’t the leader since I didn’t think I was fit. Father was calm when he replied, but I could hear the underlying fury in his tone. “ ” 0ii is none other than your subordinate. Your birth marks you as a leader, and your prowess doesn’t matter. She is below you; she always will be. You were born a light that even others cannot but bow to. You are the one who will herald humanity into the new age. You are 0i. Do not forget. I never brought it up again after that. We knew that Father loved us and Father’s word was law. So, we let go of sentimental things and focused on being the best, making Father proud. Oiii was next. Another girl. She wasn’t as charming or as beautiful as 0ii, but what she lacked in those aspects, she made up for it in smartness. She had hooked teeth and self-esteem that caved with the slightest criticism. Father called her a fond name: 𝐵𝑎𝑏𝑦 𝑅𝑎𝑏𝑏𝑖𝑡. She was the only one who had a name in five of us. When I asked him why, Father laughed heartily and spoke to me in the most jovial tone ever. “ ” What is she? Isn’t that what she looks like? I’m only calling her as I see it. Father was wise beyond his years; I admired him so much. We all did. He knew so much about everything, much more than we did. 0iv was the bookworm in our midst. He consumed all the books in the library, but still, Father paid him no heed. With his glasses perked on the bridge of his nose, 0iv did everything to get Father’s attention, but it was to no avail. I asked Father about it once again. “ ” Those who seek attention are like dogs begging for scraps. They’re unworthy of it. That’s why they demand it from others. I was the only one whom Father talked to so openly, so I spoke for us. We were the golden children humanity had been waiting for; Father said so. But when we were 9, Father introduced the blades. They were beautiful blades with edges sharp enough to sear through a human. “ ” These are the gifts I’ll be giving all of you for your 9th birthday. Use them well. Father didn’t explain anything to any of us, but I understood. We had read stories of assassins, and we understood what was required of us. Father was all-knowing, training us in the Art of the blade even though he was never there to see. It didn’t matter; 0v excelled at it beautifully. He was the only one out of us. Father never said anything bad about him. Truthfully, Father never said anything about him at all. It was as though he didn’t exist. 0v never complained, either. He was even more beautiful than 0ii, yet his gaze was even chillier than Father’s. The day 0iii died, Father didn’t bat an eyelid. 0v was playing with two sets of blades, and 0iii dared to approach him in that mode. I always knew he was darker than he let on, but watching 0iii grip her bloodied neck, eyes widened in disbelief as she dropped to the floor in a thud made me finally understand. Father was passing by when it happened, but he didn’t spare her a glance. The rest of us were shocked by what we’d seen, but 0v kept twirling the blades in the air, a bored expression on his beautiful face. “ ” I screamed for the first time at Father. But Father didn’t look as though he was bothered. He walked away. Father didn’t look back for the first time. Father! 0iii is dead! Once we were five, then we became four. Father stopped talking to me after that. 0v acted normally, while 0ii and 0iv steered clear of him. 0v offered no explanations for his actions, and one day, Father stopped coming to see us. He usually went away for days, leaving his underlings with us sometimes. But this time, Father left us alone. I counted the days. I was the leader, and I had to tell the others what to do. It wasn’t exactly a good experience. But strangely, 0v didn’t show any sign of distress. It was as though he didn’t have emotions. “ ” You must do your share of the work, or we’ll starve. I decided to face him one day, already livid. 0iii was dead, and he couldn’t even act remorseful. He didn’t care, plain and simple. He grew up with us yet grew apart from us. It was like he was in his universe, unperturbed by everything around him. He was human yet acted so differently, I wondered, too. “ ” Starve? Okay, starve. That was all he said as he stood up from where he sat, his hair shielding his eyes. His hair had been an afro since I’d known him, adding to the mystery his personality created. I was tired of being scared of a person who was just like me. It was irksome and unnecessary. We weren’t made of fear. I placed my arms on his shoulders to pull him back, but the next thing I saw was a dagger heading straight in the middle of my face. It was too fast for me to dodge, and I saw my life flash before my eyes. His wrist movements were perfect, with no single flaw in the swing of his blade. It was funny how I was thinking about that while facing my inevitable death. As soon as it began, it ended. I was alive. But I could hear gurgling sounds as my eyes I once closed opened to see someone in front of me. It was 0iv, and the hilt of a dagger was protruding from his mouth. “ .” I said, starve 0v drew his dagger out of the open mouth of 0iv as 0iv fell to the ground, much like 0iii did. The dagger reached the back of his throat as 0v drew it out, placing his tongue on the flat of the dagger, licking the blood on it while staring at me with the coldest eyes I’d ever seen. I tried to stop my body from shaking, but it was hard. 0iv did what he did to shield me, and he lay there, choking on his blood. 0v just walked away as 0ii ran to the toilet to puke. I just stood and stared. Hell had been unleashed in the form of a beautiful boy. Once we were five, then we became three. Father returned a week after that incident, and he said nothing about the disappearance of 0iv. I didn’t explain either. I needed to find a way out of there with 0ii. Father tried to talk to me again, but it was no use. 0v was a demon, one he let roam. Father wasn’t to be trusted. I spoke hushedly with 0ii, not letting 0v into our conversation. But he didn’t mind, and it was as though he wasn’t even there. He kept fiddling with daggers, paying us no heed. That night was meant to be the night of our escape. I prepared 0ii, and in the dead of the night, we tried to sneak out. I regretted it. It was all fine and good, and we were making progress. Since everybody was asleep, we moved stealthily. 0iii would have loved it; she was the smartest of us. I still had no idea why 0iv protected me; he was book-smart. He knew the implications of playing the hero. Yet, he died for me. If I had one thing to say, I wish I’d died with them. 0ii needed to be protected at all costs. We made it past the door of Father, and at that moment, I was elated. We could finally leave. Until the lights came on. They were blinding, and I held on to 0ii, fearing the worst. “ ” Finally, let the game begin. I heard the voice of Father as he smiled sinisterly at me, 0v coming from right behind him, holding a dagger and staring at me with lifeless eyes. Then, hell began. Protecting 0ii was hard, and shielding myself from stab wounds was harder. I could hear booming sounds from the speakers that miraculously materialized around me. They were making bets on us. 0v came after me, swinging the dagger dangerously close to my face, but I evaded. Then he went after 0ii, but she showed some skill by dodging his attacks. Even in the cramped space of the facility, we fought for our lives. I saw Father approaching me, and 0v stopped swinging his dagger and stood still. “ ” I tried to reach him, but he slapped my hands away. It felt even worse than losing 0iii and 0v. Father… “ ” Don’t you dare call me that. I’m not your father, and I’ll never be the father of freaks. Here, we harvest your abilities and sell them to the world. You’re not special; you’ve never been. You never had a father, just a childish imagination and daddy issues. I couldn’t speak. I didn’t know what happened, but I could hear ringing in my ears and a voice screaming at me from afar; it was muffled as though it was coming from a distance. The ringing sound in my ears got louder as 0ii screams came out in muffled gasps, but I was too far away to hear. Far away in my head. I could listen to the sound of a blade rending through flesh; I could smell the sickening odor of blood. But, I was rooted at that spot. Nothing could make me move as I stared with despair at the man I called Father. “ ” I’m sorry, I’m so sorry, 0i. You always were the best of us. I was too weak; I believed them. 0v eyes were filled with tears as he turned the blood-stained dagger on himself. Even then, I didn’t react. Even when his screams filled the room, and the ringing in my ears faded like background noise. Once we were five, then we became one. I was left alive. Not like I could die. They didn’t kill me; they just made me live every day in remembrance of those I killed. We weren’t the golden children of humanity, we were livestock. Not all is as it seems. But hell is on earth. When I called him Father, I unleashed it. I didn’t die. I didn’t live either. I was in oblivion. It was truly damnation.

Stories, News & Mentions about AAIR

Aair’s Stories on HackerNoon

Aair's News & Mentions