ComSL: A Composite Speech-Language Model for End-to-End Speech-to-Text Translation

Joint speech-language training is challenging due to the large demand for training data and GPU consumption, as well as the modality gap between speech and language. We present ComSL, a speech-language model built atop a composite architecture of public pretrained speech-only and language-only models and optimized data-efficiently for spoken language tasks. Particularly, we propose to incorporate cross-modality learning into transfer learning and conduct them simultaneously for downstream tasks in a multi-task learning manner. Our approach has demonstrated effectiveness in end-to-end speech-to-text translation tasks, achieving a new state-of-the-art average BLEU score of 31.5 on the multilingual speech to English text translation task for 21 languages, as measured on the public CoVoST2 evaluation set.

Paper

Similar papers

Peer review

Reviewer nRjc6/10 · confidence 5/52023-06-28

Summary

Authors propose a speech-text pretrained model for spoken language tasks which leverages already existing pre-trained speech and language models. Modality mapping/alignment is based on a concatenation of paired speech-text (no need of word level alignment). The model is evaluated on a speech-to-text translation (S2TT) task (on CoVoST2 dataset) and slightly outperform previous models such as Whisper for S2TT.

Strengths

-Good performance reported on CoVoST speech-to-text translation tasks -An approach that leverage existing pre-trained models (whisper speech encoder; mBART text2text) to build efficient ST systems

Weaknesses

-Difference/positioning related to previous speech-text pretrained models should be improved: in what aspect does your method really differ from SpeechT5, SLAM, etc ? -Improvement over Whisper Large when model is trained without pseudo ST data is tiny (and maybe not significant) so does the improvement really comes from the new approach/losses proposed or does it come only from the use of pseudo ST data ? -Only Whisper speechencoder is used, why ? Authors could have also chosen wav2vec multilingual (XLS-R) or HuBERT speech encoders

Questions

-In what aspect does your method really differ from SpeechT5, SLAM, etc ? -"Speech Transformer blocks are initialized with the encoder parameters of Whisper model" why using Whisper speech encoder only and not XLS-R (multilingual wav2vec) or HuBERT ? -"Different from the previous speech-text alignment approaches that rely on externally forced-alignment methods to determine word or other unit boundaries in speech sequences," => Do SpeechT5 & SLAM need external word-level forced-alignment (i don't think so) -(tab 1) does the improvement really comes from the new approach/losses proposed or does it come only from the use of pseudo ST data ?

Rating

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

3 good

Presentation

3 good

Contribution

2 fair

Limitations

Limitation mentioned are the use of mBART encoder decoder instead of large decoder only models (with bigger architectures)

Reviewer 7Di56/10 · confidence 4/52023-07-06

Summary

This paper proposes a composite speech-language model for speech-to-text (ComSL) translation. ComSL first leverages existing pre-trained models for initialization, including Whisper speech recognition model and mBART machine translation model. And then proposed several modality alignment methods that do not require additional tools. Extensive experiments on CoVoST2 demonstrate the superiority of the proposed method, which also achieves a new SOTA performance.

Strengths

1. This paper provides insight into designing large-scale spoken language models. 2. This paper proposes several methods based on concatenated embeddings to align the representation of speech and text. 3. The proposed method outperforms previous work and achieves a new SOTA performance.

Weaknesses

1. The training process may be unstable. As shown in the appendix, multiple losses are distributed by various weights from 0.1 to 0.8. The authors do not explain how these hyperparameters are determined, and how much the impact is. A solid system is expected. 2. Although this paper achieves a new SOTA performance, it mainly relies on the combination of several common techniques. Therefore, the contribution is limited. To achieve SOAT performance, ComSL is trained with pseudo ST data, resulting in unfair comparisons.

Questions

See weaknesses

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

Not applicable

Reviewer KhJu6/10 · confidence 5/52023-07-07

Summary

This work presents a speech-language model built from both speech-only and language-only pretrained models. By compositing pre-trained models from 2 modalities, the authors show that a data-efficiency for spoken language tasks can be achieved. In particular, the authors proposed a few cross-modality loss functions which are designed to build a strong relationship between 2 modalities. The authors demonstrated that their method is able to achieve a new sota BLEU on the CoVoST2 evaluation task.

Strengths

This paper is well written. It explains the motivations and methods clearly. Besides great overall results, it also presents a detailed analysis including ablation study and in-detail examination of usage of the cross-modality.

Weaknesses

- It is a bit disappointing to see from the result the cross modality tasks, which is the main contribution in this work, does not give significant difference in terms of BLEU score (Table 2, 29.40 --> 29.69 due to CML loss). I believe the follow up improvement from pseudo data can be obtained without CML loss. - It would be better if the authors can clearly state how many parameters are in the pre-trained ASR encoder, the LM (mBART) and in the adapters; and which parts are finetuned or fixed during training. - Though the model is trained with both ASR and AST loss, the authors only evaluate the model on the AST / MT tasks. It would be great if the author can provide a benchmark on ASR as well. - It is not immediately clear to me that whether the comparison of this work to USM / whisper is fair. It seems to be both whisper and USM does not finetune on CoVoST2 tasks, while they both use CoVoST2 tasks as an out-of-domain evaluation. On the other hand, whisper is able to perform both ASR and AST using the same model with just different prompt, while the authors did not reveal their ASR performance. There are also a recent paper relevant to this work: Rubenstein, Paul K., et al. "AudioPaLM: A Large Language Model That Can Speak and Listen." arXiv preprint arXiv:2306.12925 (2023). Understandably, this paper appears after NeuralPS submission deadline. The authors may consider to cite it since it achieves a new SOTA on CoVoST2.

Questions

- After the finetuning on CoVoST2, I think the model can still perform ASR ? But it would be great that the authors to clearly state that (or explain why it cannot).

Rating

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

3 good

Presentation

3 good

Contribution

2 fair

Limitations

I think the authors have addressed limitations adequately.

Reviewer tyJC6/10 · confidence 4/52023-07-08

Summary

With the goal to improve speech translation task, this work, leverages a multi-task training approach optimizing weighted sum of ASR, ST, MT and cross-modality learning (CML) objectives. Using mBART encoder and decoder Transformer blocks, the CML objective is purposed to better align speech and text modalities. CML involves a masked token prediction (MTP), speech to text mapping (STM), and encoder representation matching (ERM) objectives. Overall, the proposed arc (ComSL) includes a speech and textual transformer blocks, and a two layered bride adapter module. Before the multi-task step, pre-trained models (mBART and Whisper) are fine-tuned using training data for 50 languages, mainly involving CoVoST2 data set. Results are provided for low, medium and high-resource languages show ComSL's performance improvements over recently proposed multi-modal (text and speech) models such as USM and Whisper.

Strengths

- This work, is well motivated suggesting an cross-modality learning objectives, to address the modality gap in the next frontier for a working composite speech and textual model. - The results for ST are encouraging, performing similarly or slightly better than recently proposed multi-task based speech-text models like Whisper and USM.

Weaknesses

- Although, ComSL is well motivated to focus on bridging the modality gap, the pre-training and fine-tuning using in-domain data doesn't seem to deliver significant improvements. With close results, and lack of significance testing, its rather difficult to interpret if the proposed approach is better performing. - Given the focus on multi-task learning, the provided evaluation (only for ST task) is unexpected. Adding at least ASR task in to the comparison could have provided more substance to the report. - Discussion and analysis part could have focused on ST cases (or examples) where CML improves over other models without using the cross-modality objectives. This would show the importance of introducing the objective.

Questions

- What was the motivation for not evaluating other tasks, as this model is trained in a multi-task setting, at least the ASR evaluation makes sense. - Given the number of objectives involved, particularly for the CML part, how does model training and inference complexity compares? - Table 1, all models ... are fine-tuned with CoVoST2 data, does this include USM model? if not, what is the justification for a fair comparison? - Table 1, for Non-E2EST, does it make more sense to use Whisper + and a standard MT model trained separately? - Given most numbers are quite close, in table 1-3, adding statistical significance could make the results more meaningful, and to the support SOTA assertion. - L103 fix unfinished sentence - S3.3 how does the split is formulated for z^s and z^x before passing it to mBART decoder? Obviously the split is necessary to optimize the MTP objective using z^x, have you considered directly using e^x without the concat step?

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

2 fair

Contribution

3 good

Limitations

Limitations are included, reflecting the main content of the work. I suggest authors to consider a separate limitation part and to consider societal impact if available.

Authorsrebuttal2023-08-16

We are delighted that our response addressed your concerns and are grateful for your insightful feedback and the updated score.

Authorsrebuttal2023-08-16

We would like to express our gratitude to the reviewers for taking the time to review our manuscript again. As we approach the end of the authors-reviewers discussion period, we welcome any remaining concerns and are happy to address them at the reviewers’ request.

Reviewer KhJu2023-08-20

Thanks very much for the detailed response and explanation. I am satisfied with the author's response.

Authorsrebuttal2023-08-20

Dear Reviewer tyJC, We appreciate the time and effort you have put into reviewing our paper. It is our hope that our rebuttal has successfully addressed all of your concerns. We would like to draw your attention to the fact that we had provided the ASR evaluation results in the appendix of the supplementary materials. We sincerely hope that this clarification will positively influence your opinion on the quality of our submission. If you have any further questions or comments, we would be more than happy to provide further clarification. Thank you! Authors

Reviewer 7Di52023-08-21

I appreciate the detailed explanations. I will update the score from 5 to 6.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC