Tell What You Hear From What You See -- Video to Audio Generation Through Text

The content of visual and audio scenes is multi-faceted such that a video can be paired with various audio and vice-versa. Thereby, in video-to-audio generation task, it is imperative to introduce steering approaches for controlling the generated audio. While Video-to-Audio generation is a well-established generative task, existing methods lack such controllability. In this work, we propose VATT, a multi-modal generative framework that takes a video and an optional text prompt as input, and generates audio and optional textual description of the audio. Such a framework has two advantages: i) Video-to-Audio generation process can be refined and controlled via text which complements the context of visual information, and ii) The model can suggest what audio to generate for the video by generating audio captions. VATT consists of two key modules: VATT Converter, a LLM that is fine-tuned for instructions and includes a projection layer that maps video features to the LLM vector space; and VATT Audio, a transformer that generates audio tokens from visual frames and from optional text prompt using iterative parallel decoding. The audio tokens are converted to a waveform by pretrained neural codec. Experiments show that when VATT is compared to existing video-to-audio generation methods in objective metrics, it achieves competitive performance when the audio caption is not provided. When the audio caption is provided as a prompt, VATT achieves even more refined performance (lowest KLD score of 1.41). Furthermore, subjective studies show that VATT Audio has been chosen as preferred generated audio than audio generated by existing methods. VATT enables controllable video-to-audio generation through text as well as suggesting text prompts for videos through audio captions, unlocking novel applications such as text-guided video-to-audio generation and video-to-audio captioning.

Paper

Similar papers

Peer review

Reviewer BS9v6/10 · confidence 5/52024-07-06

Summary

This paper proposes a multimodal video-to-audio generative model called **VATT** that can generate audio and audio captions given input silent videos and optional text prompts. It is capable of performing controllable text-guide video-to-audio generation and video-to-audio captioning. The framework consists of two parts: i) **VATT converter**, an instruction fine-tuned LLM that projects video features into language space. To train such a model, the authors caption the existing audio-visual dataset by LTU-13B. ii) **VATT audio**, a bi-directional transformer audio decoder based on **Maskgit** to generate audio tokens and text tokens with given inputs by iterative parallel decoding and Encodec as audio tokenizer. Through the quantitative and qualitative experiments on the VGGSound dataset, they demonstrate that the proposed method achieves state-of-the-art performance in the aspect of audio quality and inference speed.

Strengths

- The paper is well-written and easy to follow. The motivation is clear. All details are clearly stated in the paper and appendix. - The paper has several technical highlights, especially on the V2A Instruction Tuning part which enables text control on the video-to-audio generation. - The authors perform the throughout experiments and ablations to demonstrate its technical advance. The proposed method outperforms baselines on the quantitative evaluation and human study.

Weaknesses

- The reviewer thinks the major contribution lies in the V2A Instruction Tuning part which enables the text control on the video-to-audio generation. While the contribution of maskgit-based audio decoder seems to be incremental. - From the reviewer's perspective, using the VATT converter to encode visual information will lose some temporal information from the video. From the provided qualitative examples, the reviewer thinks the proposed method is not performing quite well on the generated synchronized audio but only roughly correct.

Questions

The reviewer has questions about the paper: - The reviewer wonders why Maskgit as the decoder framework. From the reviewer's side, the inference speed seems to be the main advantage. - For the visual encoder eva-clip, it encode the feature frame-by-frame. The reviewer thinks it might lose temporal information, do authors explore this with a different encoder? - The reported inference speed of Diff-Foley is much slower than the number reported in their paper, any clue why it is happening

Rating

6

Confidence

5

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes, the authors have adequately addressed the limitations.

Reviewer Gn3y5/10 · confidence 4/52024-07-09

Summary

The paper introduces a two-stage model for video-to-audio generation controlled by text. In the first stage, a large language model (LLM) generates audio captions based on the video content. In the second stage, the model uses video frames and the generated audio captions to predict masked audio tokens, thereby generating the audio.

Strengths

1. An anonymous website is provided to showcase the audio generation results, clearly illustrating the effectiveness of the model. 2. The proposed method can accomplish both video-to-audio generation and captioning tasks, and it demonstrates superior performance compared to the mentioned audio generation models.

Weaknesses

1. The content of the paper is densely packed, with many details relegated to the appendix, making it somewhat difficult to understand. 2. There are some grammatical errors in the paper, such as in lines 133-134 where it states "showed improvement in multi-modal understanding in in tasks such as captioning and question-answering." 3. The paper's novelty is somewhat limited. I noticed that the bidirectional-attention, Iterative Parallel Decoding, and masking design mentioned in the paper were already proposed in reference [1]. How do these differ from what is presented in this paper? If the only contribution is integrating video input into a large language model, I believe this approach alone is insufficient to support the acceptance of the paper. [1] Zalán Borsos, Matt Sharifi, Damien Vincent, Eugene Kharitonov, Neil Zeghidour, and Marco Tagliasacchi. Soundstorm: Efficient parallel audio generation. arXiv preprint arXiv:2305.09636, 2023.

Questions

1. See the 3rd item in “Weakness”. 2. In the quantitative evaluation of video-to-audio captioning, comparing the proposed model with LLAVA, which only accepts single image input, seems unfair. Why not compare it with large language models designed for video understanding, such as VideoChat or Video-LLaMA?

Rating

5

Confidence

4

Soundness

2

Presentation

2

Contribution

2

Limitations

Lack of novelty, too dense content

Reviewer rzvn5/10 · confidence 5/52024-07-12

Summary

In this paper, the authors propose VATT, a multi-modal generative framework for text-guided video-to-audio generation. VATT consists of two key modules VATT Converter and VATT Audio. The former component maps video features to LLM vector space with a projection layer. The latter one generates audio tokens with a bi-directional transformer. Experiments on datasets VGGSound and AudioSet-2M demonstrate that the proposed VATT framework surpasses existing SOTA in terms of KLD score.

Strengths

1. The authors identified limitations within the existing video-to-audio generative modeling techniques, indicating a clear understanding of the challenges in the field. 2. The publication of the generated samples is also commendable.

Weaknesses

1. As stated by the authors, the diversity of text generated by LLM could influence the generated audio. However, it is unclear how the quality of the generated caption affects the results. It is suggested that the authors include an ablation study comparing the results of different captions quality and choose some examples to analyze the reasons. 2. In the section on quantitative results, the proposed method is only evaluated on VGGSound, which is not enough to prove the generalization and robustness of the proposed method. It would be more convincing to conduct experiments and ablation studies on more datasets. 3. There seems to be a writing error in part D of Appendix, “we use existing an existing audio large language model…”

Questions

See weaknesses

Rating

5

Confidence

5

Soundness

2

Presentation

3

Contribution

2

Limitations

yes

Authorsrebuttal2024-08-12

Dear Reviewer, We are very thankful for your valuable feedbacks and comments! Let me know if you have further questions regarding the rebuttal response or any other questions related to the paper. We look forward to your further feedbacks! Thanks, Authors of Paper Submission 1915

Authorsrebuttal2024-08-13

Since the discussion period ends soon, we want to reach out to see if you have any further advice or feedback and whether there is more information we could provide for reconsideration of your score? Please let us know and we will be happy to engage in further discussions. Thanks!

Reviewer 5yrj6/10 · confidence 4/52024-07-13

Summary

- The paper proposes a model for video-to-audio generation (main task) and video-to-audio captioning (auxilliary task). - The text is optionally used to control audio generation for ambiguous video cases. - They use two step approach i.e. video-to-caption stage and video+text-to-audio stage - Stage 1: video-to-caption - This stage basically converts videos and text to audio relevant captions - They train a VATT converter that uses LoRA finetuning to get audio captions. - This stage uses AudioSet and VGGSound and captions generated using LTU model. - Stage 2: video+text-to-audio - This stage generates audio tokens, given video+text as input. The video+text is converted to audio relevant captions using stage-1 - They use LM based audio token generation. The VATT Audio decoder learns to predict the masked audio tokens. During inference, iterating parallel decoding is used. - VGGSound is used to train this stage. - There model shows comparative and better performance on video-to-audio(V2A) generation task and text to audio generation task. Add text to V2A task further improves KLD score. - Stage 1 when compared to baselines can perform better on video-to-audio captioning. - Qualitative and human evaluation shows preference towards their work.

Strengths

- The paper proposes a new problem of utilizing text for video-to-audio generation. - Several novel and interesting technical contributions. i.e. - Video+text conditioning to generate audio by converting this condition to joint audio-caption. - Using synthetic captions for training and LoRA for MLLM finetuning. - Using iterative parallel decoding and bi-directional self-attention. - Their approach improves on video-to-audio generation, and text-to-audio generation both quantitatively and qualitatively. - The results also show improvement in video-to-caption generation.

Weaknesses

- Text captions generated by LTU are used as ground truth for video-to-audio generation, text-to-audio generation, and video-to-audio captioning. The authors should verify the correctness of the captions. - Table 2 results for Text-to-audio generation seem slightly unfair for the baselines. - The LTU-generated caption may have some biases or patterns of caption generation that VATT has seen during training while the baselines have not. - Though AudioGen has seen VGGSound tags, it has not seen the captions generated by LTU (which VATT has). - AudioLDM2 has not even seen VGGSound data for training. - A fair comparison would be finetuning/training from scratch Text-to-audio models on VGGSound with LTU-generated captions. Or VATT should be compared on AudioCaps (with videos downloaded from YouTube). - Table 3 results are also a little unfair: - Similar to the argument above LLAVA has not seen LTU-generated captions for VGGSOUND. - Also LLAVA takes only 1 frame while VATT takes more frames. - Maybe a fair comparison would be to train some captioning models on LTU-generated data. Minor comments: - The captions for the tables should come above the table. For eg. Table 1,2,3 - Missing reference: A previous work by Mo. et al [1] that utilizes text+video-to-audio generation. This should be mentioned - Align-acc for diff-foley is 82.47 in the paper vs 94.05 in the original paper [1] DiffAVA: Personalized text-to-audio generation with visual alignment

Questions

My questions are based on the arguments mentioned in the weakness section: - Can the correctness of the LTU-generated captions be verified? Since they are used as the ground-truth for several tasks? - Can the baseline text-to-audio(TTA) generation models be trained on VGGSound+LTU captions? Can VATT be compared on AudioCaps? - Can VATT be compared on other video captioning datasets? Or the existing video-captioning models be finetuned/trained from scratch on this LTU data? - Why are some metrics different from the numbers quoted in the original paper? For eg. Diff-foley alignment score.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes the authors adequately addressed the limitations.

Reviewer Gn3y2024-08-10

The author's response basically solved my confusion, so I chose to boost my score to borderline accept.

Authorsrebuttal2024-08-10

Thank you so much for your valuable feedback and time devotion for reviewing the work along with the rebuttal! We will incorporate these modifications in our revised manuscript.

Reviewer BS9v2024-08-10

Thanks for the authors' response. The reviewer's concerns are mostly addressed. I'd like to keep my positive score.

Authorsrebuttal2024-08-10

Thank you again for your insightful comments, suggestions, and appreciation of our work! We will adjust our manuscript to reflect all the changes made in the rebuttal.

Reviewer 5yrj2024-08-10

Thanks for the comprehensive explanation of my questions. Hence, I would keep my positive score. Please include the additional experiments and clarifications in the main text or appendix of the paper.

Authorsrebuttal2024-08-10

Thank you for your detailed and thoughtful reviews, and appreciation of our work! We will include all the mentioned items in the rebuttal into our revised manuscript properly.

Authorsrebuttal2024-08-11

We thank the reviewer for providing feedback regarding use of data in our submission. The benchmarks for audio-visual data for training and testing of models used in our work are from VGGSound [1] and AudioSet [2]. These are standard open creative commons and are datasets and benchmarks for audio-visual learning tasks as well as audio generation tasks. For example, in the video-to-audio generation literature, prior state of the art works such as Diff-Foley [3], FoleyGen [4], SpecVQGAN [5] and Im2wav [6] use VGGSound for training and testing. Similarly, in audio-visual understanding literature, CAV-MAE [7] utilizes both VGGSound and AudioSet-2M for training and testing on audio-video retrieval tasks. Therefore, it is a standard practice to use these two open and publicly released datasets and to compare novel proposed approaches performance on the same datasets with these prior works in the field. Similar to prior works, we extracted audio and visual features from these videos in the aforementioned datasets using deep neural networks from publicly released pretrained vision and audio backbones for training and testing our models and other baselines. We do not keep the visual frames and the audio waveforms besides a few samples which are used for demonstration purposes, i.e., qualitative samples to evaluate qualitatively the approach performance. These few samples are not publicly available for demonstration purposes and are provided exclusively via anonymous link (webpage) for reviewers for confidential review. Upon publication of our work we will make sure that the published samples in the manuscript or on the public webpage have explicit permission by creators for the publication. Besides samples, upon publication of our work we do not intend to publish data besides synthetic text we created for training and testing of our model and the pretrained model for research purposes. We will consult and obtain permission from our institution/s and follow NeurIPS guidelines if this becomes relevant. We also wanted to note that in our work we performed a user study and included the results of the study in the manuscript. We followed our institutional process for determination whether IRB approval is applicable, and upon determination that it is, we have applied and obtained such approval from our institution. We will follow the reviewer recommendation and state that explicitly in the final version of the manuscript. Sincerely, Authors of Paper Submission 1915 References: [1] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 721–725. IEEE, 2020. [2] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 776–780. IEEE, 2017. [3] Simian Luo, Chuanhao Yan, Chenxu Hu, and Hang Zhao. Diff-foley: Synchronized video-to-audio synthesis with latent diffusion models. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors, Advances in Neural Information Processing Systems, volume 36, pages 48855–48876. Curran Associates, Inc., 2023. [4] Xinhao Mei, Varun Nagaraja, Gael Le Lan, Zhaoheng Ni, Ernie Chang, Yangyang Shi, and Vikas Chandra. Foleygen: Visually-guided audio generation. arXiv preprint arXiv:2309.10537, 2023. [5] Vladimir Iashin and Esa Rahtu. Taming visually guided sound generation. arXiv preprint arXiv:2110.08791, 2021. [6] Roy Sheffer and Yossi Adi. I hear your true colors: Image guided audio generation. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE, 2023. [7] Gong, Y., Rouditchenko, A., Liu, A. H., Harwath, D., Karlinsky, L., Kuehne, H., and Glass, J. Contrastive audio-visual masked autoencoder. arXiv preprint arXiv:2210.07839, 2022b.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC