SpeechAlign: Aligning Speech Generation to Human Preferences

Speech language models have significantly advanced in generating realistic speech, with neural codec language models standing out. However, the integration of human feedback to align speech outputs to human preferences is often neglected. This paper addresses this gap by first analyzing the distribution gap in codec language models, highlighting how it leads to discrepancies between the training and inference phases, which negatively affects performance. Then we explore leveraging learning from human feedback to bridge the distribution gap. We introduce SpeechAlign, an iterative self-improvement strategy that aligns speech language models to human preferences. SpeechAlign involves constructing a preference codec dataset contrasting golden codec tokens against synthetic tokens, followed by preference optimization to improve the codec language model. This cycle of improvement is carried out iteratively to steadily convert weak models to strong ones. Through both subjective and objective evaluations, we show that SpeechAlign can bridge the distribution gap and facilitating continuous self-improvement of the speech language model. Moreover, SpeechAlign exhibits robust generalization capabilities and works for smaller models. Code and models will be available at https://github.com/0nutation/SpeechGPT.

Paper

References (22)

Scroll for more · 10 remaining

Similar papers

Peer review

Reviewer BxMB8/10 · confidence 4/52024-06-25

Summary

This paper proposes to apply preference optimization techniques (which have proven useful in aligning text language models’ outputs to users’ preferences) to speech language models that generate sequences of discrete audio representations and then speech. The particularity of the preference dataset is that it does not rely on truly human preferences but is simply made up of audio (AR) tokens obtained from natural speech and synthetic speech (the ones coming from natural speech being the preferred ones). Different preference optimization (PO) strategies are investigated: RLHF-PPO, DPO, etc and human evaluation is conducted to evaluate the synthetic speeches obtained. DPO seems to be the most performant PO method. An iterative process where the updated speech generation model is used to create a new and more challenging preference dataset (with synthetic / natural speech pairs) further improve the speech generation quality.

Strengths

The integration of human feedback to align speech outputs to human preferences is a new topic addressed in this paper with convincing results on speech generation.

Weaknesses

-the preliminary analysis on ‘distribution gap’ is interesting in itself but it is not clear how it really relates with SpeechAlign (in other words: why SpeechAlign solves this distribution gap observed earlier) -the way Best-of-N Sampling (BoN) is presented is confusing: it is at the same level of PO methods but while reading its description it looks more like a decoding approach than a model alignement approach => ??

Questions

-4.1 Exp setup: how the training parameters (lr, bs) are chosen ? -Few is said about the vocoder that generates speech from the acoustic tokens ? -Will the PO datasets be shared ?

Rating

8

Confidence

4

Soundness

4

Presentation

4

Contribution

4

Limitations

-I don't see so many limitations to this paper

Reviewer cBrC8/10 · confidence 3/52024-07-09

Summary

This paper introduces a method to improve speech generation in a speech language model via preference optimization. The method relies on creating a dataset of "gold" speech tokens produced by a neural codec model from a speech sample, contrasted with synthetic tokens produced by a speech generating model from text. The model is then trained to preferably generate tokens closer to the "gold" ones, via a number of preference optimization methods, and evaluated using both automatic metrics (WER and speaker similarity) and human preference judgments. The results show that the model can iteratively improve its performance based on the preference dataset.

Strengths

The method is simple yet non-obvious. The existence and impact of the distribution mismatch between gold and synthetic tokens is demonstrated via preliminary analysis and experiments. A number of different preference optimization algorithms are tested. The evaluation is carried out on two datasets, and accompanied by a thorough analysis and several ablations.

Weaknesses

The description of the preference optimization algorithms (COH, DPO, RLHF-PPO) is hard to understand without already being familiar with the relevant papers. The results are presented without showing their variance, even though the underlying data should be available as the paper mentions evaluating each model 10 times. It would be good to have the spread of the scores available in addition to the mean.

Questions

What is meant by the "reference model" in line 173?

Rating

8

Confidence

3

Soundness

4

Presentation

3

Contribution

4

Limitations

No discussion of limitations within the body of the paper.

Reviewer VNUv5/10 · confidence 4/52024-07-11

Summary

This study analyzes the training-inference mismatch that occurs in codec language models, a branch of personalized speech synthesis research, and mitigates it through preference optimization methods. By avoiding the labor-intensive process of collecting human preference test results, the researchers efficiently gathered data and used it to further fine-tune the model, achieving improved results in personalization.

Strengths

1. Preference optimization has not yet been sufficiently explored in personalized speech synthesis, and this study demonstrates its effectiveness for this purpose. 2. They conducted thorough observations and analyses of the problem they defined. 3. The illustrations and figures added to the paper aid in comprehension, and the writing is clear and easy to understand. 4. They evaluated various components and details they used within the paper, effectively demonstrating the impact of each component.

Weaknesses

I believe the evaluation with other baselines might be insufficient. I have included additional evaluation questions below for consideration.

Questions

1. (Questions) Distribution Gap in Section 2.3: - You mentioned a distribution gap between golden AR tokens and synthetic AR tokens. Which attributes cause this distribution difference? Generally, discrete features are known to be more robust against error propagation compared to continuous features, and many studies have leveraged this characteristic for TTS. Does the distribution gap arise because the SpeechTokenizer fails to remove acoustic information within the AR token, leaving residual information that causes a mismatch in synthetic AR tokens? What specific information is being mismatched? 2. Scalability and Dataset Size: - While recent zero-shot personalized speech synthesis models typically train on large-scale datasets of at least tens of thousands of hours, this study used the relatively smaller LibriSpeech dataset. This raises questions about the scalability of your methodology. Can the model achieve better performance with more data, as demonstrated by many recent models that show excellent personalized speech synthesis performance using vast datasets? Evaluating the model's potential with larger datasets is crucial for understanding its full capabilities. Although Section 5.2 provides an analysis of data size, it is limited to the amount of data from LibriSpeech and still falls short compared to recent studies in terms of data volume and data distribution (data from various source such as GigaSpeech, ...). 3. Baselines - Given the smaller dataset size compared to recent zero-shot approaches, it would be helpful to know the relative competitiveness of your current model (trained solely on LibriSpeech) against recent personalization technologies. Even if the performance falls short compared to models trained on over 10,000 hours of data, if the gap is not significant, this could indicate an important direction of research focused on efficient personalized modeling with less data. Although recent personalized speech synthesis trends show a lack of open-source models, it might still be possible to use samples from their demo pages to compare metrics like SIM, QMOS, and SMOS.

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

They addressed their limitations in the appendix.

Reviewer VNUv2024-08-07

Thank you for your kind response.

Thank you for your kind response. Regarding Q2, I realize I may have misunderstood the point. Nonetheless, I appreciate the experiments you conducted to increase the data. Additionally, thank you for the comparison with current high-performing models in response to Q3. I will raise the score to 5 points. I have an additional question: Is your methodology applicable to non-autoregressive state-of-the-art models (e.g., NaturalSpeech, Voicebox, SoundStorm, HierSpeech++, etc.)? If so, have you attempted to apply it to any of these models?

Authorsrebuttal2024-08-07

Response to additional question

Thank you for improving the scores! Regarding your new question, I believe that the SpeechAlign method can also be applied to non-autoregressive speech generation methods. For discrete-token-based methods (such as SoundStorm), we can use the same process as in SpeechAlign to construct a preference dataset and then modify the cross entropy loss in SoundStorm to the DPO loss. For diffusion/flow-matching-based methods (such as Naturalspeech2 and Voicebox), we can refer to Tango 2[1] and SPIN-Diffusion[2], which apply DPO to diffusion. We'll expand SpeechAlign to non-autoregressive methods in the next version. [1] Majumder, Navonil, et al. "Tango 2: Aligning diffusion-based text-to-audio generations through direct preference optimization." arXiv preprint arXiv:2404.09956 (2024). [2] Yuan, Huizhuo, et al. "Self-play fine-tuning of diffusion models for text-to-image generation." arXiv preprint arXiv:2402.10210 (2024).

Reviewer ZtfR7/10 · confidence 4/52024-07-22

Summary

The paper introduces "SpeechAlign," a method aimed at improving text-to-speech (TTS) performance by aligning speech generation with human preferences. It addresses the distribution mismatch between ground truth AR tokens and predicted AR tokens in neural codec language models. The proposed method involves preference optimization and iterative training, which has shown to enhance the performance of TTS systems.

Strengths

* Insightful Analysis: The paper provides a valuable insight into the distribution mismatch between ground truth AR tokens and predicted AR tokens, identifying it as a key issue affecting TTS performance. * Effective Methodology: The proposed preference optimization and iterative training strategies are well-reasoned and demonstrate a clear improvement in TTS performance. * Comprehensive Evaluation: The experimental results are extensive and include both subjective and objective evaluations, providing strong evidence for the effectiveness of the proposed method.

Weaknesses

* Performance Gap: Despite the improvements, the performance of the proposed method still falls short of the ground truth. * Limited Scope: The method is specifically tailored to the VALL-E based TTS model and does not present a general framework for other types of speech language models.

Questions

* Hyperparameter Tuning: Did the authors comprehensively tune the hyperparameters of the baseline methods, including exploring different learning rates? * Reward Model Accuracy: What is the accuracy of the reward model used in the preference optimization process? Iterative Training: Why does only DPO have iterative training, while SpeechAlign-RLHF-PPO does not? What justifies this choice? * Generalizability: Can the method generalize to other single-stage codec-based TTS systems like Voicecraft, where there is no distribution mismatch between AR and NAR models, and a single AR model generates all levels of speech tokens?

Rating

7

Confidence

4

Soundness

2

Presentation

3

Contribution

3

Limitations

* The method is more like a specific method to alleviate the domain mismatch or error propagation of the pipeline codec-based TTS model. * The method relies on ground truth tokens as the chosen samples in the preference data, potentially overlooking other high-quality token sequences that can also serve as the chosen samples.

Area Chair GyyM2024-08-12

Hello, Reviewer. The author has submitted a response to your comments. Whether or not it addresses your concerns, it would be greatly appreciated if you could acknowledge that you have reviewed the reply.

Reviewer ZtfR2024-08-13

Reviewer response by Reviewer ZtfR

Thanks for the response. The author address most of the questions, and the follow up work [1] on Voicecraft does convince me that the method can be generalized to other codec-based TTS. I raised my score from 6 to 7.

Reviewer cBrC2024-08-14

Many thanks for your response. My positive assessment of the paper remains unchanged.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC