Pengi: An Audio Language Model for Audio Tasks

In the domain of audio processing, Transfer Learning has facilitated the rise of Self-Supervised Learning and Zero-Shot Learning techniques. These approaches have led to the development of versatile models capable of tackling a wide array of tasks, while delivering state-of-the-art performance. However, current models inherently lack the capacity to produce the requisite language for open-ended tasks, such as Audio Captioning or Audio Question&Answering. We introduce Pengi, a novel Audio Language Model that leverages Transfer Learning by framing all audio tasks as text-generation tasks. It takes as input, an audio recording, and text, and generates free-form text as output. The input audio is represented as a sequence of continuous embeddings by an audio encoder. A text encoder does the same for the corresponding text input. Both sequences are combined as a prefix to prompt a pre-trained frozen language model. The unified architecture of Pengi enables open-ended tasks and close-ended tasks without any additional fine-tuning or task-specific extensions. When evaluated on 22 downstream tasks, our approach yields state-of-the-art performance in several of them. Our results show that connecting language models with audio models is a major step towards general-purpose audio understanding

Paper

References (68)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer 4yQP5/10 · confidence 4/52023-07-06

Summary

Pengi is a novel audio language model that leverages transfer learning to frame all audio tasks as text-generation tasks. It takes an audio recording and text as input, and generates free-form text as output. The input audio is represented as a sequence of continuous embeddings by an audio encoder, and the text input is represented similarly by a text encoder. Both sequences are then combined as a prefix to prompt a pre-trained frozen language model. When evaluated on 22 downstream tasks, Pengi achieved state-of-the-art performance in several of them.

Strengths

1. The proposed model is a novel audio language model that can be used for multiple audio tasks, including close-ended and open-ended tasks. It does not require any additional fine-tuning or task-specific extensions. 2. The paper introduce a new learning framework that frames all audio tasks as audio and text input to text output tasks. This framework uses a single training procedure and a captioning objective function. For training, Pengi uses new audio task templates inspired by Instruction Tuning. 3. Pengi has been extensively evaluated on 21 (or 22?) downstream tasks across various audio domains. It achieved state-of-the-art performance in several of these tasks, establishing a baseline for general-purpose ALM.

Weaknesses

.

Questions

1. Please add more baseline - a cascade model for each task. 2. Does the model has capabilities of ASR? 3. Can you add ablation analysis of each component?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

2 fair

Contribution

4 excellent

Limitations

.

Reviewer sUJx5/10 · confidence 5/52023-07-06

Summary

This paper explores Large Language Models (LLMs) in the context of audio processing. The core idea is to do audio-injected instruction tuning for pre-trained LLM. The method is simple: collecting a lot of audio-text paired data and use it to fine-tune a pre-trained CLAP audio encoder together with a frozen LLM. Audio is injected by concatenating the CLAP feature sequence before text. Taking advantage on the nature of LLMs training strategy, the pair relation can be provided in any form (e.g., captioning, label, metadata) and simply put into a pre-defined template to train with next token prediction. Evaluation is done on a wide collection of audio-related tasks using the output sentence of the LLM given the input audio and question.

Strengths

- This is the first (together with few concurrent papers) work to explore LLM for audio processing. - Performance on audio captioning is solid. - The method is simple in a good way, and this work can be served as a baseline for audio processing with LLMs if the evaluation can be more completed (see weaknesses).

Weaknesses

The biggest concern I have for this paper is that **the evaluation seems to be limited/biased**. - Table 3: In Pengi's framework, CLAP serves as a feature extractor that is pre-trained and unfrozen. It is obvious that Pengi can (and should) be making improvement over CLAP considering the scale of the system, the total amount of data used, and the computational power required. Comparing Pengi to state-of-the-art methods on each bench mark would better justify its value. (Imho, this is also why LLM is a great success - they generalized to and performed well on most of the tasks so good to the point where people can live with the amount of data/computation it costs.) Even if the numbers are not overwhelming, it will still add value to this work as a first step of exploring LLM's application in audio. - Table6/Section5.4: It is unclear how the authors obtained the retrieval performance for existing contrastive methods using their text-to-text retrieval pipeline. How do they index the dataset with text given the audio/text encoder from contrastive learning? Moreover, the text-to-text setup could be biased and unfair for other methods, since Pengi builds on top of pre-trained LLM and focuses on text during training. While a 100% fair comparison might not be possible, it is clearly unfair to use significantly more resource AND compare in a way favoring the large model. Table 5/7 are good examples where at least the evaluation protocol is fair and I don't see a reason why Table 6 should be treated differently. Again, even if the numbers are not good enough, it would still help the community by establishing a standard for general purpose audio model. Overall, this work is interesting in terms of exploring general audio processing model, but the evaluation should be improved. I would happily raise my score if the above-mentioned concerns can be addressed.

Questions

- The use of text encoder: line 112 said it's fundamental but there seems to be no ablation on this. Having LLMs taking text as input seems to be more intuitive. Any idea why? - For line 270: "... by 32%", isn't this 22-ish%?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

2 fair

Presentation

4 excellent

Contribution

2 fair

Limitations

- Generalizability (data): Table 8 shows that in the cases (namely, ESC50, US8K, and DCASE17) where Pengi have not seen the dataset, results are not as competitive. It would be good if more details on evaluation pipeline can be provided and interesting to see some error analysis. - Generalizability (format): All the tasks this paper considered are covered by the templates hence the input is not in free form. One of the key strength of instruction-tuned LLM is the ability to take any input and answer accordingly. are there examples where the question (text input) is never seen during training?

Reviewer vWxh7/10 · confidence 5/52023-07-09

Summary

This work is inspired by the visual language models (VLM) in the literature and presents an audio language model (ALM) for various audio tasks, including both open-ended and close-ended tasks. It also presents a comprehensive evaluation of the proposed ALM on a range of both open-ended and close-ended tasks, and showed very promising results.

Strengths

This paper presents an innovative audio language model which leverage a pretrained audio encoder and a pretrained language model. Though the proposed network architecture is largely borrowed from vision language models like [47], it is still relatively new to the audio domain. This paper is well-written and presents comprehensive evaluations on 22 downstream tasks.

Weaknesses

I feel this work is somewhat limited in that it only uses a relatively small language model (GPT-2 line, 124M params) and there is no study of how the language model can affect the performance on the downstream tasks. It is possible that with a stronger language model, the performance of the proposed model on open-ended tasks can be further improved, while the performance on close-ended tasks may more depending on the audio encoder quality.

Questions

- I feel it is a little bit counter-intuitive that in Figure 2, a `m2` mapping network is used after the text encoder, as text encoder in Figure 2 is already encoded the text prompt into the same space as the text in the response part. What's the purpose of `m2` here and do you have experimental results to show using `m2` is helpful ? - It is also unclear to me how the network handle the variable length in text prompt. It looks like to me that both audio prompts and text prompts in this works has a fixed length, i.e., 40 tokens for each part. However, it is unclear from the paper, how the authors would handle variable lengths of text prompt. It is also unclear what the effect of this fixed length (40 in the experiments) on the audio task performance. - The authors have proposed 2 methods for evaluating the proposed model's performance on close-ended tasks: log-likelihood and text matching. However, all the following experiments using text matching and there is no comparison of log-likelihood based method vs text matching methods. Are they give relatively close results?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

The authors have addressed the limitations of the proposed methods adequately.

Reviewer o28f4/10 · confidence 4/52023-07-10

Summary

This paper proposes a new audio-language learning model by treating existing audio tasks as text-generation tasks. The model architecture allows both open-ended and close-ended tasks. By evaluating on 22 downstream tasks, this paper shows competitive performance on many of them.

Strengths

The idea of treating various forms of audio tasks as text generation task is reasonable, which allows scaling the size of the data to large models. The evaluation is pretty extensive. Authors compared the proposed model on multiple benchmarks against multiple models. The paper is well-written and easy to follow.

Weaknesses

It seems that the strongest baseline is LAION CLAP [53], while authors did not compare with it in Table 3 or 9. In Table 8, Pengi did not outperform LAION, which seems to indicate that this model does not outperform the SOTA approach. My other concern is that the model's performance is largely based on the LLM and thus limited to their weaknesses. This has been discussed in the limitation section as well.

Questions

My main concern about this work is its performance against SOTA. I'd appreciate authors' clarification on that. ==== post rebuttal === I'd like to thank the authors for their responses. However, the performance improvement against LAION is not really convincing and comparisons are lacking, and thus I'm keeping the original score.

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

It has been discussed in Section 6.

Reviewer iz2N6/10 · confidence 4/52023-07-27

Summary

The authors present an approach to combine multiple non-ASR audio tasks into a single model. Motivated by audio language models and VLM, the authors propose using a frozen LLM to create an audio LM that can be used for open-ended (purely generative) and close-ended (classification) tasks. The main idea is to use an audio encoder and a text encoder to create fixed length prompts that can be used with a frozen LM after training. The authors train various tasks by using task specific text prompts (“generate audio caption”, “this is the sound of”, etc.). In this sense, the model is similar to a multitask learning model (more on this below). Results show that the model is competitive compared to other approaches in the literature that are task-specific, and an alternative multi-task model, CLAP.

Strengths

- A single model that can handle multiple tasks, both open-ended and close-ended. - One of the first approaches that incorporate VLM / LLMs to create a general purpose audio LM tasks. - Results show gains using LLM for building the model, compared to a baseline that doesn't.

Weaknesses

- While the model can cover a number of tasks with competitive results, the model still performs worse on some of the tasks considered compared to task-specific models. - Furthermore, it is unclear if the method generalizes to new tasks, which is a main strength of non-audio LLMs / VLMs. The authors show zero-shot capabilities for new labels within an existing task. Although impressive, not entirely novel given prior works have approached the problem along similar veins (like CLAP, e.g.) but in limited settings. The results in Tab. 9 also show poor performance in zero-shot settings for a new task. - The results in Tab. 8 also show that the model is not as best as some of the other models on zero-shot classification task.

Questions

1) The authors tokenize and map text prompts. Why is an explicit mapping needed? Can’t they re-use the same tokenizer as the LLM? 2) Why does the audio and text prompt prefix have to be of fixed length? Does the length of the input audio affect quality? 3) Line 180 and Fig. 3: Text matching is not appropriately described and is a little hard to follow. Consider providing a more detailed description. Are the embeddings created for each token or for the entire sentence? If for the sentence, how are they created (summarized)? 4) The authors claim sota in a few different close-ended tasks. Are the gains coming from the new formulation (audio LM) or the extra datasets that the authors are now using? For example, what if the authors train a single model with multiple classification heads for the N-tasks (close-ended, at least)? Does it work worse than the current model that uses the frozen LLM and the proposed architecture? 5) Line 181: When computing log likelihoods, do the authors normalize based on the number of tokens in the desired values (dog barking vs. sea, e.g.)? Does the number of tokens affect the overall score? 6) Sec. 5.5: Does the presented result show that perhaps a good portion of the performance comes from training data. For instance, wav2vec2 works better on emotion recognition, likely because the training data is relevant. Are the training set the same for the remaining non-speech models in Tab. 9? Minor 1) The authors use close-ended and open-ended tasks a lot, but only define it Sec. 4.2. Consider moving this to an earlier section to avoid confusion. 2) The authors say the text encoder is frozen (Line 111). Where does this encoder come from? 3) Why are the results in Tab. 4 for the best setting, worse than those in Tab. 7? 5) Please provide more descriptive captions for Tab. 4 – 6. 6) Line 271: Why does Pengi outperform human performance by such a large margin? It is interesting, so perhaps an explanation is warranted. 7) It would be useful to include the best results from Mei et al to Tab. 8. Also, it’d be useful to include some representative contrastive methods to Tab. 6 for audio retrieval. 8) Why does Pengi work better on some zero-shot tasks and not the others, compared to techniques like CLAP? Is it related to the training data? Typos: 1) Line 51: For example, -> Examples include 2) Line 70: Missing citation (?) 3) Figure 2 caption: a text a prompt -> a text prompt 4) Line 151: question: question -> question: {question} 5) Line 205: 320 secs / 1024 secs: Perhaps the authors mean milliseconds? 6) Line 270: Missing Table# (Tab. 8?).

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

1) Does not support ASR, which is, arguably, one of the most important audio task. 2) Unclear if the model can learn new tasks. The existing tasks use a fixed prefix prompt, which is tokenized and mapped. This makes it cumbersome to add new tasks to the model. This is a significant drawback compared to current LLMs that can learn new tasks by prompting. 3) How much does the LLM help? What is the quality if the LM component is trained from scratch with the current dataset?

Reviewer iz2N2023-08-14

Thank you for addressing the comments. The results using CLAP trained using the same data is especially interesting. Clarification regarding tokenizer and mapping: Wouldn't using the same tokenizer as the LLM put the tokens in the same space at the LLM? Why does it need additional mapping to "understand" the tokens?

Authorsrebuttal2023-08-14

We sincerely thank the reviewer for carefully reading our paper and rebuttal. For any additional clarifications, we are more than happy to address them. *** **Q1: Wouldn't using the same tokenizer as the LLM put the tokens in the same space at the LLM?**\ Yes, using the same tokenizer as the LLM will put them in the same space as LLM. However, our text encoder is not the same as the LLM (GPT2), so it's not producing tokens in the same space. **Q2: Why does it need additional mapping to "understand" the tokens?**\ Additional mapping is needed to understand the tokens for the case where we use a text encoder different than LLM (GPT2). If we remove the text encoder, additional mapping is not needed to understand tokens. Instead, the additional mapper helps the LLM produces different text output for different text prompts. For example, "this is sound of" should produce "dog barking" and "this emotion is" should produce "happy". This is out-of-domain for the frozen LLM and therefore requires a way to adapt the LLM for our data. We choose prefix-tuning as our choice of adaptation and introduce mapper m2. There are other ways to achieve this functionality i.e. instead of using mapper m2, one can use LORA updates or gated cross-attention on LLMs. We choose to keep the LLM completely frozen and instead tune the prefix using a mapper m2.

Reviewer sUJx2023-08-17

I would like to thank the authors for answering the questions. Overall, I am satisfied with the extra detail provided that complements this work. I will increase my rating to borderline accept.

Area Chair 3iD92023-08-18

Please read rebuttals

Dear reviewers, if you didn't already, please read the rebuttals ASAP and at least acknowledge them explicitly. Best, Area Chair

Reviewer vWxh2023-08-20

Thanks for very much for the detailed explanation!

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC