DiffNorm: Self-Supervised Normalization for Non-autoregressive Speech-to-speech Translation

Non-autoregressive Transformers (NATs) are recently applied in direct speech-to-speech translation systems, which convert speech across different languages without intermediate text data. Although NATs generate high-quality outputs and offer faster inference than autoregressive models, they tend to produce incoherent and repetitive results due to complex data distribution (e.g., acoustic and linguistic variations in speech). In this work, we introduce DiffNorm, a diffusion-based normalization strategy that simplifies data distributions for training NAT models. After training with a self-supervised noise estimation objective, DiffNorm constructs normalized target data by denoising synthetically corrupted speech features. Additionally, we propose to regularize NATs with classifier-free guidance, improving model robustness and translation quality by randomly dropping out source information during training. Our strategies result in a notable improvement of about +7 ASR-BLEU for English-Spanish (En-Es) and +2 ASR-BLEU for English-French (En-Fr) translations on the CVSS benchmark, while attaining over 14x speedup for En-Es and 5x speedup for En-Fr translations compared to autoregressive baselines.

Paper

Similar papers

Peer review

Reviewer DRdm5/10 · confidence 3/52024-07-11

Summary

This work builds on TRANSPEECH (Huang et al., 2023) by applying the diffusion method to reduce noise, thereby normalizing speech units for further generation. The authors further use classifier-free guidance to enhance non-autoregressive generation. They conduct experiments on CVSS En-Fr and En-Es datasets, comparing their methods with the baseline. Both proposed methods show improvements.

Strengths

1. This work proposes a diffusion method to normalize target speech units, which outperforms previous approaches. 2. It explores classifier-free guidance to improve NAT generation for the speech-to-speech translation task.

Weaknesses

1. The paper overstates its contributions. The authors claim to be the first to apply diffusion in the speech-to-speech translation task. However, they merely use it to generate auxiliary training targets and still follow the S2UT strategy (Lee et al., 2022b). This diminishes the novelty of the paper. 2. Although the work is based on TRANSPEECH, the authors compare only two of the three translation directions. One of the core contributions is applying the diffusion method to normalize speech units, yet the method shows minimal improvement (only 0.3 BLEU) compared to BiP (Bilateral Perturbation) in the En-Fr task. Given the method significantly outperforms others in the En-De task, this result is perplexing. 3. For a minor suggestion, in Tables 3 and 4, ensure consistent significant figures. Additionally, it would be more appropriate to place Table 3 before Table 4.

Questions

1. How does the performance fare on the CVSS Fr-En task? 2. In my opinion, the goal of normalizing the unit is to disentangle linguistic information from noisy speech features. If you use DDPM for this process, the target $z_0$, which is the output of the VAE encoder, should be the linguistic feature. How do you ensure that $z_0$ represents the desired linguistic information?

Rating

5

Confidence

3

Soundness

2

Presentation

3

Contribution

2

Limitations

The authors appropriately state the limitations and broader impacts of this work.

Authorsrebuttal2024-08-12

Invitation for Comments and Clarifications

Dear Reviewer DRdm, We greatly value your feedback and have provided clarifiications to your questions and additional experiments on Fr-En task. To ensure that we have properly addressed your concerns, we would greatly appreciate if you could review our responses and provide any further comments. We are looking forward to engaging with you before the discussion period ends. Thank you for your time and consideration.

Reviewer AuR57/10 · confidence 4/52024-07-13

Summary

This paper proposes DiffNorm, a diffusion-based self-supervised method for speech data normalization, aiming to alleviate multimodal problem in non-autoregressive speech-to-speech translation (NAT). DiffNorm consists of a VAE to reconstruct the speech feature and a diffusion model to add and remove noise of latent vector. Experiment shows that DiffNorm significantly improve NAT translation quality compared to baselines.

Strengths

1. It is surprising to see pure data normalization improves NAT so much. ASR-BLEU improves 7 BLEU on En-Es direction with DiffNorm. 2. This technique has wide applicability. All NAT S2S unit-based models can potentially benefit from it. Not to mention it requires no handcrafting rules for noise injection. 3. Ablation studies on noise level and training of DiffNorm further provide users a general guide on how to adapt DiffNorm to their own dataset and model.

Weaknesses

1. The multi-modal problem has two aspects: semantic and acoustic. DiffNorm seems able to reduce acoustic modalities. Unclear if DiffNorm can also do that on semantic modalities, i.e., multiple feasible translations for the same source input. 2. Classifier-free guidance combined with DiffNorm leads to worse performance than DiffNorm alone in Figure 4 and the authors ignore it in the text. Elaboration is needed here. 3. Lack baseline comparison with several latest S2ST models like TransSentence [1], PolyVoice [2], SeamlessM4T and etc. [1] TranSentence: speech-to-speech Translation via Language-Agnostic Sentence-Level Speech Encoding without Language-Parallel Data. [2] PolyVoice: Language Models for Speech to Speech Translation [3] SeamlessM4T—Massively Multilingual & Multimodal Machine Translation

Questions

1. Issues listed in weakness. 2. What is the role of VAE here? Can we drop it and directly use a diffusion model here for denoising? 3. Table 3 mentions larger dimension of latent vector brings better performance, why not use the original number of dimensions instead of compressing it? 4. Why improvement on En-Fr is not that significant compared to En-Es? How does language direction impact the performance? 5. Is there way to visualize the noise added on the latent vectors? Like what does it mean for the audio?

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Besides what I have already mentioned in the weakness, experiments in the paper are only conducted on En-X, but not reverse. Also, it would be interesting to see how DiffNorm works on non-European languages like Chinese and Japanese.

Authorsrebuttal2024-08-06

Rebuttal by Authors (Cont'd)

This comment follows from **Rebuttal by Authors**. > Why improvement on En-Fr is not that significant compared to En-Es? How does language direction impact the performance? > Our method has demonstrated effective performance on both language pairs, with a notable 2 BLEU point improvement for En-Fr, which is considered substantial. However, the more dramatic improvement observed in the En-Es pair can likely be attributed to specific characteristics of the dataset. Three key factors may contribute to this significant enhancement: 1. **Dataset Size**: The En-Es dataset is smaller, which means that the impact of normalization on the dataset is more pronounced, leading to more noticeable improvements. 2. **Speech Length**: Spanish speech is, on average, longer than French speech, as indicated in Table 1. Longer sequences could be more susceptible to acoustic multimodality, which our normalization method effectively mitigates. 3. **Background Noise**: The Spanish speeches from the CommonVoice dataset contain background noise. Our proposed normalization technique is capable of removing such noise, thereby enhancing the clarity and quality of the speech data. > Is there way to visualize the noise added on the latent vectors? Like what does it mean for the audio? > For visualization, please refer to Figure 5, which displays the log-mel spectrogram of the reconstructed speech. The extent of corruption in the speech varies based on the level of noise injection. As the noise injection level changes, the speech feature can be quantized into speech units that produce sounds ranging from completely unrecognizable noise to those closely resembling the original audio. > experiments in the paper are only conducted on En-X, but not reverse. Also, it would be interesting to see how DiffNorm works on non-European languages like Chinese and Japanese. > To address the concern regarding the X-En translation direction, we have conducted additional experiments for the Fr-En pair. We invite you to refer to our general response to all reviewers, where the results are detailed in the accompanying table. We observe consistent improvements in the Fr-En direction with DiffNorm, which we hope will further convince reviewers of the effectiveness of our proposed strategy. We agree that extending our method to non-European languages would be intriguing, and our self-supervised approach should be readily applicable to monolingual Chinese or Japanese speech features. However, we are currently unaware of a suitable speech-to-speech translation (S2ST) dataset for English-Chinese or English-Japanese pairs. If you could point us to such datasets, we would be eager to include them in our future studies! ------ References: [1] Lee et al., (2022). Textless speech-to-speech translation on real data. [2] Huang et al., (2023). TranSpeech: Speech-to-Speech Translation With Bilateral Perturbation [3] Ghazvininejad et al., (2019). Mask-Predict: Parallel Decoding of Conditional Masked Language Models [4] Gu et al., (2018). Non-Autoregressive Neural Machine Translation

Reviewer AuR52024-08-11

Thanks for the author response

I will keep my score.

Reviewer g4Z76/10 · confidence 4/52024-07-13

Summary

The authors introduce a process aiming to simplify the target distribution of speech-to-speech translation. This process uses a VAE model to map features to a latent space, followed by a diffusion model to normalize the features in the latent space. The authors use the generated dataset to train a non-autoregressive CMLM model on the CVSS-C dataset to validate its effectiveness.

Strengths

1. The authors propose a novel speech normalization method using Denoising Diffusion Probabilistic Models. 2. The quality gain in En-Es direction is impressive.

Weaknesses

1. I am confused by the rationale behind using a diffusion model to normalize the speech representation. The authors first add noise to the speech representation (Forward Process) and then remove the noise (Backward Process). This design seems awkward to me. Why do you think this process of adding and then removing noise can help with normalization? 2. Using a Variational Autoencoder to map features to a latent space seems contradictory to the motivation of reducing data multimodality. The VAE provides an indefinite mapping, which may hinder efforts to reduce multimodality. Additionally, $z_0 = f(h; \theta_{enc})$ is not correct; the VAE provides a distribution over the latent space. A more suitable expression would be $z_0 \sim p(z; f(h; \theta_{enc}))$. 3. The experiments are only conducted on synthesised dataset. As the paper's contribution is to alleviate the multi-modality problem in data, conducting experiments on a real speech-to-speech dataset, like [1], is much better to support the major claims. 4. The baseline results (Conformer in EnEs) seem quite low compared to the reported results in the literatre. [1] Lee et al., Textless speech-to-speech translation on real data.

Questions

1. A simple CG method in En-Es can bring an improvement about 4.5 BLEU (Table 2, Line 7), any reasons behind it? 2. In Table 4, Acc-Rec, BL-Rec, and BL-Dn seem highly correlated when applying your method. BL-Dn performs well with minor perturbation and vice versa. However, when there is no perturbation, BL-Dn degrades dramatically. What is the reason for this? Is there a borderline value of $T$ that separates these two regions? I think this exploration is important for understanding the effectiveness of using the diffusion process to normalize speech.

Rating

6

Confidence

4

Soundness

2

Presentation

3

Contribution

3

Limitations

The authors have addressed the limitations in Appendix E.

Authorsrebuttal2024-08-06

Rebuttal by Authors (Cont'd)

This comment follows from **Rebuttal by Authors**. > A simple CG method in En-Es can bring an improvement about 4.5 BLEU (Table 2, Line 7), any reasons behind it? > We are also intrigued by the significant improvement from CG on the En-Es dataset. We suspect this may be due to two factors: (1) the relatively small size of the En-Es dataset, which contains fewer than 80k data points for training, and (2) the longer sequence lengths of Spanish speech, averaging 256 tokens in the training set and 308 in the test set, which worsen the acoustic multimodality issue. CG operates by constraining the model to learn exclusively from the distribution of the target (Spanish) speech units, leading to a more consistent distribution for translation during inference. **This is particularly beneficial for longer target sequences.** To further support our claim, we recently extended our exploration of CG to English-German translation using the WMT14 dataset. Following the same settings as the CMLM model [5] and incorporating our proposed CG for the MT task, we observed that while the CMLM+CG model does not show significant improvement using the full WMT14 dataset, it does enhance performance by approximately 1 BLEU point when the training data is filtered to only include sequences longer than 60 tokens and the test data to only include sequences longer than 30 tokens, as demonstrated below: | Model | T=5 | T=10 | T=15 | |-------------------|------|------|-------| | CMLM | 19 | 20.2 | 20.5 | | CMLM+CG | 19.8 | 20.8 | 21.04 | For speech-to-speech translation that has much longer unit length, the effect of CG will be more noticeable, and therefore achieving large improvement over the baseline. > ….when there is no perturbation, BL-Dn degrades dramatically. What is the reason for this? Is there a borderline value of 𝑇 that separates these two regions? > We believe it is dataset-dependent to judge whether there is a borderline value of T that has drastic improvement. We performed more ablation studies using T=10 and T=30 as start time for noise injection and the result is shown below: | Start Time | Acc-Rec | BL-Dn | | --- | --- | --- | | T=10 | 93.8 | 15.98 | | T=30 | 91.6 | 17.29 | As indicated in the result table, with a very small T=10, the reconstruction accuracy is approximately 0.94, leading to a downstream ASR-BLEU score of 15.98. This represents a significant improvement compared to the baseline CMLM result of 12.58. The accuracy of 0.94 suggests that around 6% of tokens differ after reconstruction, which is likely due to the regularization effect of the VAE model, as T is too small for the diffusion model to cause reconstruction errors. This implies that the Spanish speech dataset may contain significant acoustic variations and noisy information, which the VAE model can already partially remove. With a larger T, the diffusion model further cleans up the representation, resulting in even higher downstream results (as observed from T=10 to T=100, where there is an increase in ASR-BLEU score). ----- References: [1] Lee et al., (2022). Textless speech-to-speech translation on real data. [2] Huang et al., (2023). TranSpeech: Speech-to-Speech Translation With Bilateral Perturbation [3]Chen et al., (2024). Robust Classification via a Single Diffusion Model [4] Duquenn et al., (2021). Multimodal and Multilingual Embeddings for Large-Scale Speech Mining [5] Ghazvininejad et al., (2019). Mask-Predict: Parallel Decoding of Conditional Masked Language Models [6] Richter et al., (2023). Speech Enhancement and Dereverberation with Diffusion-based Generative Models

Authorsrebuttal2024-08-12

Invitation for Comments and Clarifications

Dear Reviewer g4Z7, We greatly value your feedback and have provided clarifiications and additional experiments and analysis. To ensure that we have properly addressed your concerns, we would greatly appreciate if you could review our responses and provide any further comments. We are looking forward to engaging with you before the discussion period ends. Thank you for your time and consideration.

Authorsrebuttal2024-08-06

General Response to All Reviewers

Thank you to all reviewers for their valuable feedback and for recognizing the innovation and effectiveness of our work. In response to reviewers AuR5 and DRdm, we conducted further experiments on the French-English translation direction to validate our method's effectiveness. The results are as follows: | Model | ASR-BLEU | |------------------|----------| | Transformer | 15.44 | | Norm Transformer | 15.81 | | Conformer | 18.02 | | CMLM | 16.95 | | CMLM + BiP | 18.03 | | CMLM + DiffNorm | 19.53 | These results are consistent with our findings for En→Es and En→Fr, where the DiffNorm dataset showed a 2.58 improvement in ASR-BLEU score compared to the CMLM baseline. Due to limited computational resources and time, we have not yet conducted CG-related experiments, but we will include these results in the future. Additionally, we plan to release our code publicly, addressing suggestions from reviewers g4Z7 and DRdm, to facilitate its use with other languages and larger datasets.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC