HAM-TTS: Hierarchical Acoustic Modeling for Token-Based Zero-Shot Text-to-Speech with Model and Data Scaling

{}^{1}

Geely, China,

{}^{2}

National Institute of Informatics, Japan,

{}^{3}

SOKENDAI, Japan,

{}^{4}

Shanghai Jiao Tong University, China
{Chunhui.Wang5,bowen.zhang3}@geely.com, zengchang@nii.ac.jp, zym.22@sjtu.edu.cn
{Yefan.Zhu, Zifeng.Cai, Jian.Zhao9, zhonglin.jiang, yong.chen}@geely.com

Abstract

Token-based text-to-speech (TTS) models have emerged as a promising avenue for generating natural and realistic speech, yet they grapple with low pronunciation accuracy, speaking style and timbre inconsistency, and a substantial need for diverse training data. In response, we introduce a novel hierarchical acoustic modeling approach complemented by a tailored data augmentation strategy and train it on the combination of real and synthetic data, scaling the data size up to 650k hours, leading to the zero-shot TTS model with 0.8B parameters. Specifically, our method incorporates a latent variable sequence containing supplementary acoustic information based on refined self-supervised learning (SSL) discrete units into the TTS model by a predictor. This significantly mitigates pronunciation errors and style mutations in synthesized speech. During training, we strategically replace and duplicate segments of the data to enhance timbre uniformity. Moreover, a pretrained few-shot voice conversion model is utilized to generate a plethora of voices with identical content yet varied timbres. This facilitates the explicit learning of utterance-level one-to-many mappings, enriching speech diversity and also ensuring consistency in timbre. Comparative experiments¹¹1Demo page: https://anonymous.4open.science/w/ham-tts/ demonstrate our model’s superiority over VALL-E in pronunciation precision and maintaining speaking style, as well as timbre continuity.

HAM-TTS: Hierarchical Acoustic Modeling for Token-Based Zero-Shot Text-to-Speech with Model and Data Scaling

Chunhui Wang¹¹1These authors contributed equally to this work. ¹, Chang Zeng¹¹1These authors contributed equally to this work. ^{2 3}, Bowen Zhang¹¹1These authors contributed equally to this work. ¹, Ziyang Ma¹¹1These authors contributed equally to this work. ⁴, Yefan Zhu¹, Zifeng Cai¹, Jian Zhao¹, Zhonglin Jiang¹, Yong Chen¹ ${}^{1}$ Geely, China, ${}^{2}$ National Institute of Informatics, Japan, ${}^{3}$ SOKENDAI, Japan, ${}^{4}$ Shanghai Jiao Tong University, China {Chunhui.Wang5,bowen.zhang3}@geely.com, zengchang@nii.ac.jp, zym.22@sjtu.edu.cn {Yefan.Zhu, Zifeng.Cai, Jian.Zhao9, zhonglin.jiang, yong.chen}@geely.com

1 Introduction

In the last decade, significant strides Goodfellow et al. (2014); Kingma and Welling (2014); Van Den Oord et al. (2017); Dinh et al. (2015); Vaswani et al. (2017); Ho et al. (2020) have been made in the advancement of deep learning and neural network technologies, enabling the text-to-speech (TTS) to evolve from the cascade manner of acoustic models Wang et al. (2017); Li et al. (2019); Kim et al. (2020); Popov et al. (2021) and vocoders van den Oord et al. (2016); Kong et al. (2020); Wang et al. (2022); Kong et al. (2021) to the fully end-to-end (E2E) style Ren et al. (2021); Kim et al. (2021); Wang et al. (2023a); Jiang et al. (2023); Tan et al. (2021). These methods are not only capable of rapidly generating high-quality speech, but also adept at synthesizing more challenging vocal expressions such as singing Lu et al. (2020); Wang et al. (2023b, d). However, most TTS systems utilize continuous acoustic features such as MFCC in the frequency domain as intermediate representations for modeling, hindering from generating high-quality speech in the zero-shot scenario of timbre due to their mixture of semantic and acoustic information and difficulty of disentanglement Zhang et al. (2023); Yang et al. (2023b).

Recently, token-based TTS Borsos et al. (2022); Wang et al. (2023a); Yang et al. (2023a); Shen et al. (2023); Wang et al. (2023c); Song et al. (2024) methods have attracted extensive attention from both academia and industry due to their potential for synthesizing high-quality speech in the zero-shot scenario. Among these, the neural audio codec Zeghidour et al. (2021); Défossez et al. (2022); Yang et al. (2023b) has demonstrated immense potential to serve as the intermediate representation for TTS modeling. For example, VALL-E Wang et al. (2023a) utilizes a large language model Radford et al. (2019); Brown et al. (2020); Touvron et al. (2023a, b) to approximate the distribution of neural audio codecs Défossez et al. (2022) and can synthesize speech that closely mimics a target speaker’s voice from a mere three-second sample. However, despite their promising capabilities, we observe that these models often struggle with maintaining accurate pronunciation and consistent speaking style as well as timbre in synthesized speech. Additionally, the substantial requirement for large and diverse training data further limits their widespread adoption.

To tackle these issues, we proposed a Hierarchical Acoustic Modeling method, namely HAM-TTS, with a tailored data augmentation strategy for the token-based TTS model Borsos et al. (2022); Wang et al. (2023a); Yang et al. (2023a). Specifically, in order to alleviate the difficulty of directly modeling the mapping from text to neural audio codec in previous studies, we incorporate a latent variable sequence (LVS) containing supplementary acoustic information based on HuBERT Hsu et al. (2021) features into the TTS model. A Text-to-LVS predictor is optimized simultaneously with TTS model. In the inference stage, the text prompt is converted to the LVS by the predictor to provide imperative acoustic information to mitigate pronunciation errors.

Unfortunately, generating LVS based on simple HuBERT features cannot revise the issue of inconsistency of speaking style in the synthesized speech due to the personalized information contained in HuBERT features, which is a distractor to the audio prompt. Therefore, we applied the K-Means Ahmed et al. (2020) clustering method to refine HuBERT features for removing personalized information such as speaking styles, enabling the TTS model to make use of the remaining acoustic information to improve pronunciation accuracy while maintaining consistent speaking style with the audio prompt throughout the entire synthesized speech.

Timbre inconsistency is another serious problem for token-based TTS systems Borsos et al. (2022). We designed a timbre consistency data augmentation strategy to train the proposed HAM-TTS system to revise it. Concretely, we randomly replace a successive segment of a training sample with a small chunk selected from other training utterances or duplicate a successive segment of a training sample while forcing the model to predict the original utterance. It enhances the timbre consistency of the synthesized speech in the zero-shot scenario.

As illustrated in Borsos et al. (2022); Wang et al. (2023a); Shen et al. (2023), token-based TTS methods require extensive training data to assign the model the ability to synthesize diverse and high-quality speech. In this paper, instead of solely using substantial real speech data for training, we utilized a pretrained UNet-based Ronneberger et al. (2015) few-shot voice conversion model to generate voices with the same content but different timbres as a supplementary dataset, enabling the model to explicitly learn one-to-many mapping knowledge, which is beneficial to improve the diversity of generated speech and the timbre consistency.

We trained many models with different configurations on a large-scale internal Chinese dataset and evaluated them on the public AISHELL1 dataset Bu et al. (2017). We rigorously compared HAM-TTS against the state-of-the-art (SOTA) VALL-E model, which served as our baseline. The results of these experiments, conducted on a substantial dataset, clearly establish the advantages of our approach over the baseline model, demonstrating the enhanced capabilities of HAM-TTS, particularly in terms of pronunciation accuracy, speaking style consistency, and timbre continuity in challenging zero-shot scenarios.

This paper is structured to provide a comprehensive overview of our research and findings. Following this introduction, some related works are introduced in Section 2. We delve into the specifics of our hierarchical acoustic modeling method in Section 3. We then present the experimental setup and results, offering a comparative analysis with current benchmarks in Section 4. The paper concludes with a summary of our contributions and a discussion on future research directions in Section 5.

2 Related Works

Although there are many studies Tokuda et al. (2013); Li and Zen (2016); Wang et al. (2017); Li et al. (2019); Ren et al. (2021); Kim et al. (2021); Wang et al. (2023a) focusing on TTS, in this section, we only briefly review some representative works about neural audio codecs and speech generative models based on them for a closer connection to our work.

2.1 Neural Audio Codec

Recent advancements in neural audio codecs, as illustrated in Zeghidour et al. (2021); Défossez et al. (2022); Yang et al. (2023b), have significantly enhanced the field of speech synthesis. These studies collectively highlight the efficiency of neural codecs in encoding and decoding audio data, offering a more compact and flexible representation compared to traditional methods.

Soundstream Zeghidour et al. (2021) introduces a novel end-to-end neural audio codec framework, demonstrating effective compression of audio signals into a discrete latent space by residual vector quantization. This advancement facilitates the generation of high-quality audio from compact representations, highlighting the codec’s versatility in various audio applications.

Encodec Défossez et al. (2022) further explores this domain, emphasizing the codec’s role in efficiently compressing audio data while maintaining quality. Its approach showcases the potential of neural codecs in handling complex audio tasks with reduced data requirements, a crucial factor in resource-constrained environments.

In our research, these insights into neural audio codecs lay the foundation for developing a robust and efficient token-based TTS model. The enhanced fidelity and efficiency of neural codecs directly inform our approach, enabling us to achieve superior speech synthesis quality, particularly in zero-shot scenarios.

Refer to caption — Figure 1: Overview of HAM-TTS. Although it builds upon VALL-E, its design including Text-HuBERT aligner and Text-to-LVS is applicable across various token-based TTS models. To enhance the ability of HAM-TTS to process semantic information, we also let codec language models predict the phoneme sequence based on the input text in the training stage.

2.2 Token-based Speech Generation Model

More and more studies Borsos et al. (2022); Wang et al. (2023a); Shen et al. (2023); Wang et al. (2023c); Song et al. (2024) are beginning to try to use neural audio codecs as intermediate representations for speech generation. These approaches highlight the growing consensus in the field regarding the effectiveness of neural codecs in handling complex tasks.

AudioLM Borsos et al. (2022) represents a significant leap in audio generation by employing a language modeling approach. It particularly stands out for its ability to generate coherent and contextually appropriate speech, attributed to its advanced use of latent vectors conditioned on inputs. This model demonstrates how the integration of neural codecs Zeghidour et al. (2021) can facilitate the production of diverse and high-quality speech.

VALL-E Wang et al. (2023a), on the other hand, capitalizes on the neural codec’s ability Défossez et al. (2022) to approximate large language models, enabling the synthesis of speech that closely mimics a target speaker’s voice from a minimal sample.

NaturalSpeech2 Shen et al. (2023) takes these concepts further by integrating a neural audio codec with additional components such as the diffusion model. Its emphasis on zero-shot synthesis capabilities and prosody highlights the model’s robustness and versatility, particularly in generating diverse speech styles and maintaining voice quality across various scenarios.

These studies collectively underscore the importance of neural codecs in speech generation and pave the way for our research. In our work, we build upon these foundations and propose a novel hierarchical acoustic modeling approach to enhance pronunciation accuracy and speaking style consistency while utilizing a data augmentation strategy and synthetic data to emphasize the timbre consistency and diversity of generated voices.

3 HAM-TTS

The introduction of the HAM-TTS model is presented in this section. As depicted in Figure 1, in addition to the phoneme conversion and audio codec encoder components originating from the existing TTS model like VALL-E, we design a predictor to directly transform the text prompt to the latent variable sequence (LVS) to incorporate supplementary acoustic information into the neural codec language model in the inference stage. The predictor is jointly optimized with the TTS model in the training stage via the supervising signal from the output of the Text-HuBERT aligner, which utilizes the cross-attention mechanism Li et al. (2023) to align the phoneme sequence and the HuBERT features refined by K-Means clustering to generate the LVS. Detailed designs of the Text-HuBERT aligner and the Text-to-LVS predictor are presented in Section 3.1. The timbre consistency data augmentation strategy is another important contribution of our work for revising the issue of timbre inconsistency in synthesized speech. It is concretely illustrated in Section 3.2. Finally, the supplementary synthetic dataset generated by the pretrained few-shot voice conversion model is elaborated in Section 3.3. Detailed configurations for models used in our experiment will be illustrated in Appendix A.1.

3.1 Hierarchical Acoustic Modeling

We observed that previous studies like AudioLM Borsos et al. (2022) and VALL-E Wang et al. (2023a) occasionally produced speech with incorrect pronunciation. This was largely due to the limitations in directly mapping text to a neural audio codec sequence without adequate acoustic information. To address this, the Text-to-LVS predictor shown in Figure 2 is proposed to generate the latent variable sequence containing the imperative acoustic information from the phoneme sequence in the inference stage, which can be formulated as,

\boldsymbol{L}^{\prime}_{1:T_{1}}=f_{pred}(\boldsymbol{X}_{1:T_{1}}),

(1)

where $\boldsymbol{X}_{1:T_{1}}$ represents the phoneme sequence with $T_{1}$ phoneme units. $f_{pred}(\cdot)$ denotes the function of the predictor’s transformation. $\boldsymbol{L}^{\prime}_{1:T_{1}}$ is the generated LVS with the same length of the phoneme sequence. The LVS is concatenated with the corresponding phoneme sequence. Following this concatenation, the combined sequence is transformed via a convolutional layer to align with the dimension required by the neural audio codec before feeding them to the codec language model. It can be represented as,

\displaystyle\boldsymbol{S}_{1:T_{1}}=\text{Conv1d}(\text{Concat}(\boldsymbol{% X}_{1:T_{1}},\boldsymbol{L}^{\prime}_{1:T_{1}})),

(2)

where $\boldsymbol{S}_{1:T_{1}}$ is the output aligning with the dimension of audio codecs.

As illustrated in Figure 1, the Text-to-LVS predictor is simultaneously optimized with the neural codec language model in the training stage via the supervising signal generated from another new module, namely Text-HuBERT aligner. The aligner consists of $N$ blocks with the same architecture as shown in Figure 3. Each block contains $M$ residual convolution networks He et al. (2015) denoted as ResNet Block in the figure, followed by a root mean square layer normalization (RMSNorm) Zhang and Sennrich (2019), and finally a multi-head attention layer Vaswani et al. (2017) is utilized to align the output sequence of RMSNorm with the HuBERT Hsu et al. (2021) features (key and value) refined by K-Means clustering. Unlike the standard layer normalization used in the Transformer model Vaswani et al. (2017), we employ RMSNorm in the aligner, enhancing its capability to handle complex sequences and achieve faster convergence. The supervising LVS with the same length of the phoneme sequence can be computed by,

\boldsymbol{L}_{1:T_{1}}=f_{aligner}(\boldsymbol{X}_{1:T_{1}},\boldsymbol{H}_{% 1:T_{2}}),

(3)

where $\boldsymbol{H}_{1:T_{2}}$ is the refined HuBERT feature sequence with $T_{2}$ length and $\boldsymbol{L}_{1:T_{1}}$ denotes the supervising LVS. $f_{aligner}(\cdot)$ means the function of Text-HuBERT aligner module. Note that it is imperative to leverage the K-Means clustering to remove personalized information from the original HuBERT feature for revising the mutation of speaking style in synthesized speech in the zero-shot scenario.

The approximation between $\boldsymbol{L}_{1:T_{1}}$ and $\boldsymbol{L}^{\prime}_{1:T_{1}}$ is measured by a L1 loss function shown as,

\mathcal{L}_{LVS}=\sum_{t=1}^{T_{1}}|\boldsymbol{L}^{\prime}_{t}-\boldsymbol{L% }_{t}|,

(4)

where $\mathcal{L}_{LVS}$ is the metric measuring how close the $\boldsymbol{L}^{\prime}_{1:T_{1}}$ is to $\boldsymbol{L}_{1:T_{1}}$ .

3.2 Timbre Consistency Data Augmentation

Timbre inconsistency of the synthesized speech has been a non-negligible problem plaguing the TTS system in the zero-shot scenario despite the fact that contemporary token-based TTS systems Wang et al. (2023a); Yang et al. (2023a) claim to enable timbre cloning. In this section, we will illustrate our proposed timbre consistency data augmentation strategy for this issue.

To ensure timbre consistency in the synthesized speech, we implemented a data augmentation strategy on our training data. Specifically, during the loading of a batch of speech data, for 10% portion, we either randomly select a continuous segment from another sample to replace a segment in the current sample, or we randomly duplicate a segment from the same sample and concatenate it to the end of that segment. In the loss calculation, neural audio codecs from samples without data augmentation are treated as ground truth for computing the cross-entropy loss with the generated codecs. This approach enables the model to develop strong resistance to timbre perturbations. Consequently, it prevents short-term timbre variations from affecting the timbre of the entire generated speech segment, thus ensuring timbre consistency in the synthesized speech.

3.3 Supplmentary Synthetic Dataset

The fact that extensive speech data are needed to train a TTS model is prohibitive for many academic researchers. For example, Audiobox Vyas et al. (2023) has scaled the size of the training data up to 100k hours, which is a heavy burden to collect that much data for academic institutions. At the same time, there are many legal risks associated with using real data without authorization. These facts motivate us to consider using synthetic data to train TTS models. In this section, we will show how to generate synthetic data as a supplementary dataset for real data.

It is difficult to collect a large amount of data for voices with single timbre and long duration in the real world, especially for more than ten seconds, which leads to sparse data for speech with long duration when training speech synthesis models and also makes it more difficult for the model to ensure the consistency of the timbre of the whole sentence when generating long speech. With this in mind, we utilize a pre-trained UNet-based Ronneberger et al. (2015) few-shot voice conversion model concretely illustrated in Appendix A.2 to generate a large amount of long speech data to compensate for the lack of real data. We randomly select 1,000 speakers with a few minutes of speech from the real data as candidates and convert around 500 hours of real speech whose duration ranges from 10 to 20 seconds in the training dataset for each candidate. Consequently, the large amount of synthetic data improves the diversity of training data by explicitly providing one-to-many mapping for the scenario of long voices, distinct from previous studies Wang et al. (2017); Ren et al. (2021); Borsos et al. (2022); Wang et al. (2023a) in which only the phoneme-level diversity was considered.

3.4 Loss Function

We follow the training strategy of VALL-E Wang et al. (2023a) regarding TTS as a conditional codec language modeling task. Two Transformer Vaswani et al. (2017) decoder-only codec language models are trained for autoregressive (AR) and non-autoregressive (NAR) modeling, respectively. We utilize the cross-entropy (CE) loss function to measure the distance between the real and the learned distribution of codecs. It can be formulated as,

\displaystyle\mathcal{L}_{codecs}

\displaystyle=\sum_{t=1}^{T_{3}}\text{CE}(\boldsymbol{A}_{t},\boldsymbol{A}^{% \prime}_{t}),

(5)

where $\boldsymbol{A}$ and $\boldsymbol{A}^{\prime}$ mean codec sequences of the ground truth and synthesized one, respectively. $T_{3}$ denotes the length of the codec sequence. $\mathcal{L}_{codecs}$ is the loss for codec generation.

Moreover, to enhance the ability of HAM-TTS to process semantic information, the teacher forcing loss is computed on the AR codec LM and the NAR codec LM to fit the distribution of input texts, and the corresponding CE loss function is shown as,

\displaystyle\mathcal{L}_{phoneme}

\displaystyle=\sum_{t=1}^{T_{1}}\text{CE}(\boldsymbol{X}_{t},\boldsymbol{X}^{% \prime}_{t}),

(6)

where $\boldsymbol{X}^{\prime}$ means the synthesized phoneme sequence. $\mathcal{L}_{phoneme}$ is the loss for text generation.

The total loss is the sum of three loss terms, illustrated as Eq. 7. More details of the training method are available in Appendix A.3.

\mathcal{L}=\mathcal{L}_{LVS}+\mathcal{L}_{phoneme}+\mathcal{L}_{codecs}

(7)

4 Experiment

Table 1: Performance comparison on AISHELL1 dataset. All models were trained exclusively on 150k hours of real data. We compare the performance of ground truth (GT), VALL-E, HAM-TTS-S, and HAM-TTS-L models, showcasing the effectiveness of HAM-TTS in pronunciation accuracy, naturalness, and speaker similarity. The NMOS, SMOS, and MOS were computed with a 95% confidence interval.

Model	#Params	CER%( $\downarrow$ )	NMOS( $\uparrow$ )	SMOS( $\uparrow$ )	MOS( $\uparrow$ )
GT	-	2.6	4.03 $\pm$ 0.08	4.30 $\pm$ 0.06	4.45 $\pm$ 0.07
VALL-E	426M	5.5	3.65 $\pm$ 0.15	4.03 $\pm$ 0.12	4.05 $\pm$ 0.10
HAM-TTS-S	421M	4.0	3.79 $\pm$ 0.11	4.12 $\pm$ 0.10	4.27 $\pm$ 0.08
HAM-TTS-L	827M	3.2	4.01 $\pm$ 0.07	4.26 $\pm$ 0.09	4.45 $\pm$ 0.07

4.1 Experiment Setup

Dataset: All TTS models were trained on our internal Chinese speech dataset comprising both real and synthetic speech. The dataset includes 150k hours of real speech and 500k hours of synthetic speech. The real speech component encompasses approximately 20,000 speakers, with each audio segment ranging between 5 to 20 seconds in length and a sampling rate of 24kHz. On the other hand, the synthetic speech dataset is derived from 1,000 speakers, with each audio segment varying from 10 to 20 seconds in length. This extensive and diverse dataset plays a critical role in the robust training and performance of our model. As for the test data, we selected 50 speakers from the public AISHELL1 dataset Bu et al. (2017) and each speaker has five sentences whose duration varies from 5-20 seconds. Since our training data has no overlap with the public dataset, all testing speakers are unseen, aiming at showing the zero-shot ability of our model.

Baseline: VALL-E Wang et al. (2023a) is used as the baseline model in our experiments since it is a representative SOTA work of token-based TTS systems. We reproduced and trained it on the internal dataset due to no official implementation available.

Evaluation metrics: We evaluate all models from three aspects: pronunciation accuracy, speaking style consistency, and timbre consistency. Pronunciation accuracy is represented by character error rate (CER) metric, which is calculated by a pretrained Whisper Radford et al. (2023) model²²2https://huggingface.co/espnet/pengcheng_aishell_asr_tra
in_asr_whisper_medium_finetune_raw_zh_whisper_multilin
gual_sp provided by ESPNet Watanabe et al. (2018). Speaking style consistency is evaluated by mean opinion score regarding the naturalness (NMOS) of speech since the mutation of speaking style is perceptible from the feedback of listeners. Timbre consistency is evaluated by the speaker similarity MOS (SMOS) metric. Additionally, we also requested all listeners to evaluate the overall quality of testing data, including the naturalness, audio quality, and pronunciation accuracy. It is represented as the MOS metric. As for the number of listeners, we employed 60 people to participate in the test. Each listener will evaluate the performance for all utterances. We believe that a listening test of this magnitude would provide a relatively objective result for the experiment.

4.2 Primary Experimental Result

In our experimental analysis, as detailed in Table 1, all models were trained exclusively on 150k hours of real data. The HAM-TTS model, designed in two variants, HAM-TTS-S and HAM-TTS-L, explores different scales of parameterization. HAM-TTS-S, matching VALL-E with 421M parameters, ensures a fair comparison, while HAM-TTS-L expands to 827M parameters, aiming to unlock the full potential of the HAM-TTS. This scaling is crucial for assessing the effectiveness of our model in various parameter configurations.

In the table, our reproduced VALL-E achieves a CER of 5.5%, an NMOS of 3.65, an SMOS of 4.03, and an overall MOS of 4.05, aligning with those presented in the original VALL-E paper Wang et al. (2023a), indicating the reliability of our experimental setup. These results demonstrate VALL-E’s proficiency in generating speech, but also highlight areas for improvement, particularly in pronunciation accuracy and naturalness compared with the result of GT. The HAM-TTS-S model achieves a CER of 4.0%, lower than VALL-E’s 5.5%, indicating better pronunciation accuracy. Its NMOS at 3.79 and SMOS at 4.12 also surpass VALL-E, suggesting improved perceived quality and speaker similarity. The HAM-TTS-L further improves these metrics, recording a CER of 3.2%, and comparable NMOS and SMOS scores to GT, illustrating the scalability and effectiveness of the HAM-TTS model in generating high-quality, realistic speech. These results demonstrate the HAM-TTS model’s superiority in pronunciation accuracy and the consistency of speaking style and timbre.

4.3 Ablation Study of K-Means

Table 2: A comparison of HAM-TTS-S model performance with and without K-Means clustering is provided to highlight the improvement in CER, NMOS, and overall MOS metrics due to K-Means feature refinement.

Model	CER%( $\downarrow$ )	NMOS( $\uparrow$ )	MOS( $\uparrow$ )
GT	2.6	4.30 $\pm$ 0.06	4.45 $\pm$ 0.09
w/o K-Means	4.2	3.63 $\pm$ 0.12	4.14 $\pm$ 0.08
HAM-TTS-S	4.0	3.79 $\pm$ 0.11	4.27 $\pm$ 0.08

In our HAM-TTS model, we employed the K-Means clustering technique to refine HuBERT features. This approach aims to remove personalized information such as speaking styles, enabling the TTS model to focus on the core acoustic information for enhancing pronunciation accuracy and maintaining consistent speaking style with the audio prompt throughout the synthesized speech.

Table 2 in our experimental results presents the effectiveness of the K-Means clustering in our model. We compared the performance of HAM-TTS-S with and without the application of K-Means clustering. The results demonstrate that the application of K-Means clustering further improves the model’s performance. Specifically, the CER for the HAM-TTS-S without K-Means clustering was 4.2%, while the implementation of K-Means clustering reduced the CER to 4.0%. This reduction in CER indicates an improvement in pronunciation accuracy, which is a direct result of the refined HuBERT features providing more accurate acoustic information.

Furthermore, the NMOS and the overall MOS also slightly improved with the use of K-Means clustering. The NMOS increased from 3.63 to 3.79, and the MOS increased from 4.14 to 4.27, indicating that the speech synthesized with the refined features was perceived as more natural and of higher quality by listeners. These results clearly illustrate the impact of K-Means clustering in enhancing the overall performance of the HAM-TTS-S, affirming its effectiveness in providing a more accurate and consistent speaking style in synthesized speech.

4.4 Ablation Study of Synthetic Data

Table 3: Experimental result to show the effectiveness of synthetic data. We trained the HAM-TTS-S model with different sizes and combinations of real(R) and synthetic(S) data.

Training data	CER%( $\downarrow$ )	SMOS( $\uparrow$ )	MOS( $\uparrow$ )
GT	2.6	4.30 $\pm$ 0.06	4.45 $\pm$ 0.07
150k(R)	4.0	4.12 $\pm$ 0.10	4.27 $\pm$ 0.08
150k(R)+150k(S)	3.6	4.26 $\pm$ 0.09	4.32 $\pm$ 0.07
150k(R)+500k(S)	2.8	4.32 $\pm$ 0.07	4.49 $\pm$ 0.08
150k(S)	4.5	4.05 $\pm$ 0.10	4.10 $\pm$ 0.13
300k(S)	4.1	4.13 $\pm$ 0.07	4.25 $\pm$ 0.08
500k(S)	3.3	4.25 $\pm$ 0.06	4.35 $\pm$ 0.06

In our HAM-TTS model, synthetic data plays a pivotal role in enhancing the diversity and quality of the generated speech. We focused on demonstrating the impact of this synthetic data through a series of experiments, the results of which are detailed in Table 3.

The experiments were conducted using the HAM-TTS-S model, trained on different combinations and sizes of real and synthetic data. Our findings clearly show the significant improvements synthetic data brings to the model’s performance. When trained solely on 150k hours of real data, the HAM-TTS-S model achieves a CER of 4.0%, an SMOS of 4.12, and an overall MOS of 4.27. However, when augmented with synthetic data, there is a marked improvement in all metrics.

Specifically, training with an additional 150k hours of synthetic data (150k(R)+150k(S)) reduces the CER to 3.6%, and further increases the SMOS to 4.26 and the MOS to 4.32. This improvement is even more pronounced when the model is trained with an additional 500k hours of synthetic data (150k(R)+500k(S)), resulting in a CER of 2.8%, an SMOS of 4.32, and an MOS of 4.49. These results clearly indicate that synthetic data not only contributes to the reduction in pronunciation errors but also significantly enhances the quality of the synthesized speech since it enables the model to explicitly learn the knowledge of utterance-level one-to-many mappings.

Furthermore, the results underscore the promise of training HAM-TTS models solely on synthetic data. When the model was trained with varying amounts of synthetic data (150k(S), 300k(S), and 500k(S)), we observed a steady improvement in all evaluation metrics, approaching the performance levels of the model trained on real data. The model trained with 500k hours of synthetic data achieved a CER of 3.3%, closely matching the 2.8% CER of the model trained with a combination of real and synthetic data. This finding is particularly promising as it suggests that high-quality TTS systems can be developed even in scenarios where access to large amounts of real speech data is limited, highlighting the potential of synthetic data in training effective speech synthesis models.

These findings illustrate the significant impact of synthetic data in improving the performance of HAM-TTS models, both when used in conjunction with real data and when used exclusively, marking a substantial advancement in the field of speech synthesis.

5 Conclusion and Future Work

In this study, we have introduced HAM-TTS, a novel text-to-speech system that leverages a hierarchical acoustic modeling approach. This system integrates advanced techniques such as K-Means clustering for refining HuBERT features and a comprehensive strategy incorporating both real and synthetic data. Our experiments demonstrate the effectiveness of HAM-TTS in improving pronunciation accuracy, speaking style consistency, and timbre consistency in zero-shot scenarios.

Despite these significant advancements, future work could explore the optimal combination of synthetic data in terms of speaker diversity and duration per speaker. This aspect could lead to further enhancements in handling a wide range of speech variations. Additionally, optimizing the inference speed of the HAM-TTS model is crucial for enhancing its practical usability, making it suitable for real-time applications and user interactions. The exploration of these avenues will contribute significantly to advancing the field of speech synthesis.

Limitation

We acknowledge that while our HAM-TTS model has demonstrated significant advancements, certain aspects remain unexplored and present opportunities for future research. One such area is the optimal combination of synthetic data in terms of speaker diversity and duration per speaker. We have not yet investigated whether a greater number of speakers with less duration per speaker or fewer speakers but more duration per speaker would be more beneficial. This aspect is crucial for enhancing the model’s ability to handle a wide range of speech variations and could potentially lead to further improvements in the model’s performance.

Another limitation is the inference speed of the HAM-TTS model. Although the model achieves high-quality speech synthesis, the current inference process is not as efficient as it could be. There is considerable room for improvement in this area, particularly in terms of reducing the time taken to generate speech. Optimizing the model’s architecture and streamlining the inference pipeline could significantly enhance the practical usability of HAM-TTS, making it more suitable for real-time applications and user interactions.

Addressing these limitations will be a focus of our future work, aiming to refine the HAM-TTS model further and expand its applicability in various speech synthesis scenarios.

Ethics Statement

This research adheres to ethical standards in AI and speech synthesis, emphasizing data privacy, consent, and inclusivity. We address the potential for bias in our datasets and ensure fairness across diverse voices. Recognizing the risks of misuse, we advocate for responsible use and transparency in our methodology. Our work aims to contribute positively to technological advancements, balancing innovation with societal and individual well-being.

References

Ahmed et al. (2020) Mohiuddin Ahmed, Raihan Seraj, and Syed Mohammed Shamsul Islam. 2020. The K-Means Algorithm: A Comprehensive Survey And Performance Evaluation. Electronics, 9(8):1295.
Borsos et al. (2022) Zalán Borsos, Raphaël Marinier, Damien Vincent, Eugene Kharitonov, Olivier Pietquin, Matthew Sharifi, Dominik Roblek, Olivier Teboul, David Grangier, Marco Tagliasacchi, and Neil Zeghidour. 2022. AudioLM: A Language Modeling Approach to Audio Generation. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 31:2523–2533.
Brown et al. (2020) Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language Models Are Few-Shot Learners. Advances in neural information processing systems, 33:1877–1901.
Bu et al. (2017) Hui Bu, Jiayu Du, Xingyu Na, Bengu Wu, and Hao Zheng. 2017. AISHELL-1: An Open-Source Mandarin Speech Corpus and A Speech Recognition Baseline. In 2017 20th conference of the oriental chapter of the international coordinating committee on speech databases and speech I/O systems and assessment (O-COCOSDA), pages 1–5. IEEE.
Défossez et al. (2022) Alexandre Défossez, Jade Copet, Gabriel Synnaeve, and Yossi Adi. 2022. High Fidelity Neural Audio Compression. arXiv preprint arXiv:2210.13438.
Dinh et al. (2015) Laurent Dinh, David Krueger, and Yoshua Bengio. 2015. NICE: Non-linear Independent Components Estimation.
Goodfellow et al. (2014) Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. Generative Adversarial Networks. Advances in neural information processing systems, 27.
He et al. (2015) Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. 2015. Deep Residual Learning for Image Recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770–778.
Ho et al. (2020) Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising Diffusion Probabilistic Models. Advances in neural information processing systems, 33:6840–6851.
Hsu et al. (2021) Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. 2021. Hubert: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:3451–3460.
Jang et al. (2021) Won Jang, Daniel Chung Yong Lim, Jaesam Yoon, Bongwan Kim, and Juntae Kim. 2021. UnivNet: A Neural Vocoder with Multi-Resolution Spectrogram Discriminators for High-Fidelity Waveform Generation. In Interspeech.
Jiang et al. (2023) Ziyue Jiang, Jinglin Liu, Yi Ren, Jinzheng He, Chen Zhang, Zhenhui Ye, Pengfei Wei, Chunfeng Wang, Xiang Yin, Zejun Ma, et al. 2023. Mega-TTS 2: Zero-Shot Text-to-Speech with Arbitrary Length Speech Prompts. arXiv preprint arXiv:2307.07218.
Kim et al. (2020) Jaehyeon Kim, Sungwon Kim, Jungil Kong, and Sungroh Yoon. 2020. Glow-TTS: A Generative Flow for Text-to-Speech via Monotonic Alignment Search. Advances in Neural Information Processing Systems, 33:8067–8077.
Kim et al. (2021) Jaehyeon Kim, Jungil Kong, and Juhee Son. 2021. Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech. In International Conference on Machine Learning, pages 5530–5540. PMLR.
Kingma and Ba (2015) Diederik Kingma and Jimmy Ba. 2015. Adam: A Method for Stochastic Optimization. In International Conference on Learning Representations (ICLR).
Kingma and Welling (2014) Diederik P. Kingma and Max Welling. 2014. Auto-Encoding Variational Bayes. In 2nd International Conference on Learning Representations, ICLR 2014.
Kong et al. (2020) Jungil Kong, Jaehyeon Kim, and Jaekyoung Bae. 2020. HiFi-GAN: Generative Adversarial Networks for Efficient and High Fidelity Speech Synthesis. Advances in Neural Information Processing Systems, 33:17022–17033.
Kong et al. (2021) Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. 2021. DiffWave: A Versatile Diffusion Model for Audio Synthesis. In 9th International Conference on Learning Representations, ICLR 2021.
Li and Zen (2016) Bo Li and Heiga Zen. 2016. Multi-Language Multi-Speaker Acoustic Modeling for LSTM-RNN Based Statistical Parametric Speech Synthesis. In Interspeech.
Li et al. (2023) Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the 40th International Conference on Machine Learning (ICML) 2023. JMLR.
Li et al. (2019) Naihan Li, Shujie Liu, Yanqing Liu, Sheng Zhao, and Ming Liu. 2019. Neural Speech Synthesis with Transformer Network. Proceedings of the AAAI Conference on Artificial Intelligence, page 6706–6713.
Loshchilov and Hutter (2017) Ilya Loshchilov and Frank Hutter. 2017. SGDR: Stochastic gradient descent with warm restarts. In International Conference on Learning Representations.
Lu et al. (2020) Peiling Lu, Jie Wu, Jian Luan, Xu Tan, and Li Jun Zhou. 2020. XiaoiceSing: A High-Quality and Integrated Singing Voice Synthesis System. In Interspeech.
Popov et al. (2021) Vadim Popov, Ivan Vovk, Vladimir Gogoryan, Tasnima Sadekova, and Mikhail Kudinov. 2021. Grad-TTS: A Diffusion Probabilistic Model for Text-to-Speech. In International Conference on Machine Learning, pages 8599–8608. PMLR.
Radford et al. (2023) Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2023. Robust Speech Recognition via Large-Scale Weak Supervision. In Proceedings of the 40th International Conference on Machine Learning. JMLR.org.
Radford et al. (2019) Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language Models Are Unsupervised Multitask Learners. OpenAI blog, 1(8):9.
Ren et al. (2021) Yi Ren, Chenxu Hu, Xu Tan, Tao Qin, Sheng Zhao, Zhou Zhao, and Tie-Yan Liu. 2021. FastSpeech 2: Fast and High-Quality End-to-End Text to Speech. In 9th International Conference on Learning Representations, ICLR 2021.
Ronneberger et al. (2015) Olaf Ronneberger, Philipp Fischer, and Thomas Brox. 2015. U-Net: Convolutional Networks for Biomedical Image Segmentation. In Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234–241. Springer.
Shen et al. (2023) Kai Shen, Zeqian Ju, Xu Tan, Yanqing Liu, Yichong Leng, Lei He, Tao Qin, Sheng Zhao, and Jiang Bian. 2023. NaturalSpeech 2: Latent Diffusion Models are Natural and Zero-Shot Speech and Singing Synthesizers. arXiv preprint arXiv:2304.09116.
Song et al. (2024) Yakun Song, Zhuo Chen, Xiaofei Wang, Ziyang Ma, and Xie Chen. 2024. Ella-v: Stable neural codec language modeling with alignment-guided sequence reordering. arXiv preprint arXiv:2401.07333.
Srivastava et al. (2014) Nitish Srivastava, Geoffrey E. Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. 2014. Dropout: A Simple Way to Prevent Neural Networks from Overfitting. J. Mach. Learn. Res., 15:1929–1958.
Tan et al. (2021) Xu Tan, Tao Qin, Frank Soong, and Tie-Yan Liu. 2021. A Survey on Neural Speech Synthesis. arXiv preprint arXiv:2106.15561.
Tokuda et al. (2013) Keiichi Tokuda, Yoshihiko Nankaku, Tomoki Toda, Heiga Zen, Junichi Yamagishi, and Keiichiro Oura. 2013. Speech Synthesis Based on Hidden Markov Models. Proceedings of the IEEE, 101(5):1234–1252.
Touvron et al. (2023a) Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023a. LLaMA: Open and Efficient Foundation Language Models. arXiv preprint arXiv:2302.13971.
Touvron et al. (2023b) Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023b. LLaMA 2: Open Foundation and Fine-Tuned Chat Models. arXiv preprint arXiv:2307.09288.
van den Oord et al. (2016) Aäron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves, Nal Kalchbrenner, Andrew W. Senior, and Koray Kavukcuoglu. 2016. WaveNet: A Generative Model for Raw Audio. In The 9th ISCA Speech Synthesis Workshop, page 125.
Van Den Oord et al. (2017) Aaron Van Den Oord, Oriol Vinyals, et al. 2017. Neural Discrete Representation Learning. Advances in neural information processing systems, 30.
Vaswani et al. (2017) Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention Is All You Need. Advances in neural information processing systems, 30.
Vyas et al. (2023) Apoorv Vyas, Bowen Shi, Matthew Le, Andros Tjandra, Yi-Chiao Wu, Baishan Guo, Jiemin Zhang, Xinyue Zhang, Robert Adkins, William Ngan, et al. 2023. Audiobox: Unified Audio Generation with Natural Language Prompts. arXiv preprint arXiv:2312.15821.
Wang et al. (2023a) Chengyi Wang, Sanyuan Chen, Yu Wu, Ziqiang Zhang, Long Zhou, Shujie Liu, Zhuo Chen, Yanqing Liu, Huaming Wang, Jinyu Li, et al. 2023a. Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers. arXiv preprint arXiv:2301.02111.
Wang et al. (2022) Chunhui Wang, Chang Zeng, Jun Chen, and Xing He. 2022. HiFi-WaveGAN: Generative Adversarial Network with Auxiliary Spectrogram-Phase Loss for High-Fidelity Singing Voice Generation. arXiv preprint arXiv:2210.12740.
Wang et al. (2023b) Chunhui Wang, Chang Zeng, and Xing He. 2023b. Xiaoicesing 2: A High-Fidelity Singing Voice Synthesizer Based on Generative Adversarial Network. In Proc. Interspeech 2023, pages 5401–5405.
Wang et al. (2023c) Jiaming Wang, Zhihao Du, Qian Chen, Yunfei Chu, Zhifu Gao, Zerui Li, Kai Hu, Xiaohuan Zhou, Jin Xu, Ziyang Ma, et al. 2023c. LauraGPT: Listen, attend, understand, and regenerate audio with GPT. arXiv preprint arXiv:2310.04673.
Wang et al. (2023d) Xintong Wang, Chang Zeng, Jun Chen, and Chunhui Wang. 2023d. Crosssinger: A Cross-Lingual Multi-Singer High-Fidelity Singing Voice Synthesizer Trained on Monolingual Singers. In 2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), pages 1–6. IEEE.
Wang et al. (2017) Yuxuan Wang, R. J. Skerry-Ryan, Daisy Stanton, Yonghui Wu, Ron J. Weiss, Navdeep Jaitly, Zongheng Yang, Ying Xiao, Z. Chen, Samy Bengio, Quoc V. Le, Yannis Agiomyrgiannakis, Robert A. J. Clark, and Rif A. Saurous. 2017. Tacotron: Towards End-to-End Speech Synthesis. In Interspeech.
Watanabe et al. (2018) Shinji Watanabe, Takaaki Hori, Shigeki Karita, Tomoki Hayashi, Jiro Nishitoba, Yuya Unno, Nelson Enrique Yalta Soplin, Jahn Heymann, Matthew Wiesner, Nanxin Chen, Adithya Renduchintala, and Tsubasa Ochiai. 2018. ESPnet: End-to-end speech processing toolkit. In Proceedings of Interspeech, pages 2207–2211.
Yang et al. (2023a) Dongchao Yang, Songxiang Liu, Rongjie Huang, Guangzhi Lei, Chao Weng, Helen Meng, and Dong Yu. 2023a. InstructTTS: Modelling Expressive TTS in Discrete Latent Space with Natural Language Style Prompt. arXiv preprint arXiv:2301.13662.
Yang et al. (2023b) Dongchao Yang, Songxiang Liu, Rongjie Huang, Jinchuan Tian, Chao Weng, and Yuexian Zou. 2023b. HiFi-Codec: Group-residual Vector quantization for High Fidelity Audio Codec. arXiv preprint arXiv:2305.02765.
Zeghidour et al. (2021) Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi. 2021. Soundstream: An End-to-End Neural Audio Codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 30:495–507.
Zhang and Sennrich (2019) Biao Zhang and Rico Sennrich. 2019. Root Mean Square Layer Normalization. Advances in Neural Information Processing Systems, 32.
Zhang et al. (2023) Dong Zhang, Shimin Li, Xin Zhang, Jun Zhan, Pengyu Wang, Yaqian Zhou, and Xipeng Qiu. 2023. SpeechGPT: Empowering Large Language Models with Intrinsic Cross-Modal Conversational Abilities. arXiv preprint arXiv:2305.11000.

Appendix A Appendix

A.1 Model Details

Table 4: Configuration of VALL-E in the experiment.

Component	Config	Value
Phoneme Conversion	Embedding Layer	1024
Audio Codec EncoderDéfossez et al. (2022)	Quantizer	8
Audio Codec EncoderDéfossez et al. (2022)	Codebook Size	1024
	Codebook Dimension	1024
Codec Language Model	Attention Block	14³³3In order to have a fair comparison with the HAM-TTS-S model, we increase the number of parameters of VALL-E to a comparable level by increasing two additional attention blocks.
	Heads	16
	Hidden Size	4096
	Dropout	0.1
	Output Affine Layer	1024

HAM-TTS is constructed based on the VALL-E framework, inheriting certain key architectural features. Similar to VALL-E, HAM-TTS incorporates two distinct Transformer decoders. These decoders are integral to the model’s design, each serving a specific purpose in the speech synthesis process.

One of the Transformer decoders in HAM-TTS is dedicated to autoregressive modeling. This decoder plays a crucial role in sequentially predicting each element of the output based on the previously generated elements, thereby capturing the temporal dependencies in the speech sequence.

The other Transformer decoder in HAM-TTS is utilized for non-autoregressive modeling. This approach allows for the parallel generation of output elements, which can significantly enhance the model’s efficiency by reducing the dependency on the sequential generation process.

Concrete configurations for VALL-E, HAM-TTS-S, and HAM-TTS-L are shown as Table 4, Table 5, and Table 6, respectively.

Table 5: Configuration of HAM-TTS-S in the experiment.

Component	Config	Value
Phoneme Conversion	Embedding Layer	1024
Audio Codec EncoderDéfossez et al. (2022)	Quantizer	8
	Codebook Size	1024
	Codebook Dimension	1024
Codec Language Model	Attention Block	12
	Heads	16
	Hidden Size	4096
	Dropout	0.1
	Output Affine Layer	1024
Text-to-LVS Predictor	Conv1D Layers	2
	Conv1D Kernel Size	3
	Dropout	0.1
	Output Affine Layer	2
Text-HuBERT Aligner	Attention Block	10
	Heads	8
	Hidden Size	4096
	Dropout	0.1
	ResNet Block	3
	Conv1D Layer	2
	Conv1D Kernel Size	3
	Output Affine Layer	2

Table 6: Configuration of HAM-TTS-L in the experiment.

Component	Config	Value
Phoneme Conversion	Embedding Layer	1024
Audio Codec EncoderDéfossez et al. (2022)	Quantizer	8
	Codebook Size	1024
	Codebook Dimension	1024
Codec Language Model	Attention Block	24
	Heads	16
	Hidden Size	4096
	Dropout	0.1
	Output Affine Layer	1024
Text-to-LVS Predictor	Conv1D Layers	2
	Conv1D Kernel Size	3
	Dropout	0.1
	Output Affine Layer	2
Text-HuBERT Aligner	Attention Block	10
	Heads	8
	Hidden Size	4096
	Dropout	0.1
	ResNet Block	3
	Conv1D Layer	2
	Conv1D Kernel Size	3
	Output Affine Layer	2

A.2 Pretrained Voice Conversion Model

We employed a UNet-based Ronneberger et al. (2015) voice conversion model illustrated in Figure 4 to generate 500k hours of speech data for training.

In this voice conversion model, the initial processing stage involves extracting HuBERT and F0 features from the input audio. These extracted features are then concatenated and fed into a ResNet module for preprocessing. The ResNet module is designed to transform and refine these features, outputting them in the dimensions of (96, T, F), where ‘T’ and ‘F’ represent the time and frequency dimensions, respectively.

This output feature is then introduced into the encoder of a UNet architecture. The encoder performs downsampling on the frequency dimension twice, resulting in an output with dimensions (384, T, F / 4). Following this, another ResNet module is employed to further refine the output of the encoder. The refined features are then passed to the decoder of the UNet.

In the decoding process, the frequency dimension undergoes two stages of upsampling. Prior to each upsampling step, speaker characteristics are integrated into the input. This integration is crucial for ensuring that the synthesized speech retains the unique attributes of the speaker’s voice. The final output from the decoder has the dimensions of (96, T, F), effectively restoring the original frequency dimension.

It is important to note that throughout the UNet architecture, the convolutional kernels used are of size (1,7). This specific kernel size aids in capturing the essential temporal and spectral characteristics of the speech signal.

The next stage involves the conversion of these processed features into the final waveform. This is achieved using a PostNet followed by a UnivNet vocoder Jang et al. (2021), which together ensure the synthesized speech is both natural-sounding and closely matches the original audio in terms of timbre and prosody.

A.3 Training Method

We followed the training strategy used in VALL-E to employ a dual training approach to optimize the performance of the HAM-TTS model in both autoregressive (AR) and non-autoregressive (NAR) modeling.

AR Training: The AR model is trained on the concatenation of the sequence $\boldsymbol{S}_{1:T_{1}}$ and the audio codec sequence $\boldsymbol{A}^{(1)}_{1:T_{3}}$ from the first quantizer of the Encodec model Défossez et al. (2022). It can be formulated as,

	$\displaystyle p(\boldsymbol{A}^{\prime(1)}\|\boldsymbol{A}^{(1)},\boldsymbol{S}% ;\theta_{AR})=$		(8)
	$\displaystyle\prod_{t=0}^{T}p(\boldsymbol{A}^{\prime(1)}_{t}\|\boldsymbol{A}^{% \prime(1)}_{<t},\boldsymbol{A}^{(1)},\boldsymbol{S};\theta_{AR})$

NAR Training: The NAR model is employed for the audio codecs from the second to the last quantizers. This model is conditioned on $\boldsymbol{S}_{1:T_{1}}$ , the acoustic prompt $\boldsymbol{A}^{(2:8)}_{1:T_{3}}$ , and the predicted acoustic tokens $\boldsymbol{A}^{(<i)}_{1:T_{3}}$ from the previous codebooks. Each training step randomly samples a quantizer $i\in[2,8]$ , and the model is trained to fit the distribution of codecs from the selected quantizer codebook. It can be formulated as,

	$\displaystyle p(\boldsymbol{A}^{\prime(2:8)}\|\boldsymbol{A},\boldsymbol{S};% \theta_{NAR})$		(9)
	$\displaystyle=\prod_{i=2}^{8}p(\boldsymbol{A}^{\prime(i)}\|\boldsymbol{A}^{% \prime(<i)},\boldsymbol{A},\boldsymbol{S};\theta_{AR})$		(10)

Both AR and NAR models were optimized using the Adam optimizer Kingma and Ba (2015), with a learning rate set at 0.03 and a warmup spanning the first 15,000 steps. After the warmup phase, the learning rate was managed using the CosineAnnealingLR scheduler Loshchilov and Hutter (2017). The training was conducted on a robust setup of 512 NVIDIA A100 80GB GPUs, and the model processed a batch size of 8k acoustic tokens. This extensive training was carried out over a total of 400k steps, leveraging the powerful computational capabilities of the A100 GPUs to efficiently handle the large batch size and extensive training steps.