BASE TTS: Lessons from building a billion-parameter Text-to-Speech model on 100K hours of data

We introduce a text-to-speech (TTS) model called BASE TTS, which stands for $\textbf{B}$ig $\textbf{A}$daptive $\textbf{S}$treamable TTS with $\textbf{E}$mergent abilities. BASE TTS is the largest TTS model to-date, trained on 100K hours of public domain speech data, achieving a new state-of-the-art in speech naturalness. It deploys a 1-billion-parameter autoregressive Transformer that converts raw texts into discrete codes ("speechcodes") followed by a convolution-based decoder which converts these speechcodes into waveforms in an incremental, streamable manner. Further, our speechcodes are built using a novel speech tokenization technique that features speaker ID disentanglement and compression with byte-pair encoding. Echoing the widely-reported "emergent abilities" of large language models when trained on increasing volume of data, we show that BASE TTS variants built with 10K+ hours and 500M+ parameters begin to demonstrate natural prosody on textually complex sentences. We design and share a specialized dataset to measure these emergent abilities for text-to-speech. We showcase state-of-the-art naturalness of BASE TTS by evaluating against baselines that include publicly available large-scale text-to-speech systems: YourTTS, Bark and TortoiseTTS. Audio samples generated by the model can be heard at https://amazon-ltts-paper.com/.

Cross-Speaker Encoding Network for Multi-Talker Speech Recognition

End-to-end multi-talker speech recognition has garnered great interest as an effective approach to directly transcribe overlapped speech from multiple speakers. Current methods typically adopt either 1) single-input multiple-output (SIMO) models with a branched encoder, or 2) single-input single-output (SISO) models based on attention-based encoder-decoder architecture with serialized output training (SOT). In this work, we propose a Cross-Speaker Encoding (CSE) network to address the limitations of SIMO models by aggregating cross-speaker representations. Furthermore, the CSE model is integrated with SOT to leverage both the advantages of SIMO and SISO while mitigating their drawbacks. To the best of our knowledge, this work represents an early effort to integrate SIMO and SISO for multi-talker speech recognition. Experiments on the two-speaker LibrispeechMix dataset show that the CES model reduces word error rate (WER) by 8% over the SIMO baseline. The CSE-SOT model reduces WER by 10% overall and by 16% on high-overlap speech compared to the SOT model.

QS-TTS: Towards Semi-Supervised Text-to-Speech Synthesis via Vector-Quantized Self-Supervised Speech Representation Learning

This paper proposes a novel semi-supervised TTS framework, QS-TTS, to improve TTS quality with lower supervised data requirements via Vector-Quantized Self-Supervised Speech Representation Learning (VQ-S3RL) utilizing more unlabeled speech audio. This framework comprises two VQ-S3R learners: first, the principal learner aims to provide a generative Multi-Stage Multi-Codebook (MSMC) VQ-S3R via the MSMC-VQ-GAN combined with the contrastive S3RL, while decoding it back to the high-quality audio; then, the associate learner further abstracts the MSMC representation into a highly-compact VQ representation through a VQ-VAE. These two generative VQ-S3R learners provide profitable speech representations and pre-trained models for TTS, significantly improving synthesis quality with the lower requirement for supervised data. QS-TTS is evaluated comprehensively under various scenarios via subjective and objective tests in experiments. The results powerfully demonstrate the superior performance of QS-TTS, winning the highest MOS over supervised or semi-supervised baseline TTS approaches, especially in low-resource scenarios. Moreover, comparing various speech representations and transfer learning methods in TTS further validates the notable improvement of the proposed VQ-S3RL to TTS, showing the best audio quality and intelligibility metrics. The trend of slower decay in the synthesis quality of QS-TTS with decreasing supervised data further highlights its lower requirements for supervised data, indicating its great potential in low-resource scenarios.

Towards High-Quality Neural TTS for Low-Resource Languages by Learning Compact Speech Representations

This paper aims to enhance low-resource TTS by reducing training data requirements using compact speech representations. A Multi-Stage Multi-Codebook (MSMC) VQ-GAN is trained to learn the representation, MSMCR, and decode it to waveforms. Subsequently, we train the multi-stage predictor to predict MSMCRs from the text for TTS synthesis. Moreover, we optimize the training strategy by leveraging more audio to learn MSMCRs better for low-resource languages. It selects audio from other languages using speaker similarity metric to augment the training set, and applies transfer learning to improve training quality. In MOS tests, the proposed system significantly outperforms FastSpeech and VITS in standard and low-resource scenarios, showing lower data requirements. The proposed training strategy effectively enhances MSMCRs on waveform reconstruction. It improves TTS performance further, which wins 77% votes in the preference test for the low-resource TTS with only 15 minutes of paired data.

A Multi-Stage Multi-Codebook VQ-VAE Approach to High-Performance Neural TTS

We propose a Multi-Stage, Multi-Codebook (MSMC) approach to high-performance neural TTS synthesis. A vector-quantized, variational autoencoder (VQ-VAE) based feature analyzer is used to encode Mel spectrograms of speech training data by down-sampling progressively in multiple stages into MSMC Representations (MSMCRs) with different time resolutions, and quantizing them with multiple VQ codebooks, respectively. Multi-stage predictors are trained to map the input text sequence to MSMCRs progressively by minimizing a combined loss of the reconstruction Mean Square Error (MSE) and "triplet loss". In synthesis, the neural vocoder converts the predicted MSMCRs into final speech waveforms. The proposed approach is trained and tested with an English TTS database of 16 hours by a female speaker. The proposed TTS achieves an MOS score of 4.41, which outperforms the baseline with an MOS of 3.62. Compact versions of the proposed TTS with much less parameters can still preserve high MOS scores. Ablation studies show that both multiple stages and multiple codebooks are effective for achieving high TTS performance.

A Multi-Scale Time-Frequency Spectrogram Discriminator for GAN-based Non-Autoregressive TTS

The generative adversarial network (GAN) has shown its outstanding capability in improving Non-Autoregressive TTS (NAR-TTS) by adversarially training it with an extra model that discriminates between the real and the generated speech. To maximize the benefits of GAN, it is crucial to find a powerful discriminator that can capture rich distinguishable information. In this paper, we propose a multi-scale time-frequency spectrogram discriminator to help NAR-TTS generate high-fidelity Mel-spectrograms. It treats the spectrogram as a 2D image to exploit the correlation among different components in the time-frequency domain. And a U-Net-based model structure is employed to discriminate at different scales to capture both coarse-grained and fine-grained information. We conduct subjective tests to evaluate the proposed approach. Both multi-scale and time-frequency discriminating bring significant improvement in the naturalness and fidelity. When combining the neural vocoder, it is shown more effective and concise than fine-tuning the vocoder. Finally, we visualize the discriminating maps to compare their difference to verify the effectiveness of multi-scale discriminating.

Improving Adversarial Waveform Generation based Singing Voice Conversion with Harmonic Signals

Adversarial waveform generation has been a popular approach as the backend of singing voice conversion (SVC) to generate high-quality singing audio. However, the instability of GAN also leads to other problems, such as pitch jitters and U/V errors. It affects the smoothness and continuity of harmonics, hence degrades the conversion quality seriously. This paper proposes to feed harmonic signals to the SVC model in advance to enhance audio generation. We extract the sine excitation from the pitch, and filter it with a linear time-varying (LTV) filter estimated by a neural network. Both these two harmonic signals are adopted as the inputs to generate the singing waveform. In our experiments, two mainstream models, MelGAN and ParallelWaveGAN, are investigated to validate the effectiveness of the proposed approach. We conduct a MOS test on clean and noisy test sets. The result shows that both signals significantly improve SVC in fidelity and timbre similarity. Besides, the case analysis further validates that this method enhances the smoothness and continuity of harmonics in the generated audio, and the filtered excitation better matches the target audio.

Phonetic Posteriorgrams based Many-to-Many Singing Voice Conversion via Adversarial Training

This paper describes an end-to-end adversarial singing voice conversion (EA-SVC) approach. It can directly generate arbitrary singing waveform by given phonetic posteriorgram (PPG) representing content, F0 representing pitch, and speaker embedding representing timbre, respectively. Proposed system is composed of three modules: generator $G$, the audio generation discriminator $D_{A}$, and the feature disentanglement discriminator $D_F$. The generator $G$ encodes the features in parallel and inversely transforms them into the target waveform. In order to make timbre conversion more stable and controllable, speaker embedding is further decomposed to the weighted sum of a group of trainable vectors representing different timbre clusters. Further, to realize more robust and accurate singing conversion, disentanglement discriminator $D_F$ is proposed to remove pitch and timbre related information that remains in the encoded PPG. Finally, a two-stage training is conducted to keep a stable and effective adversarial training process. Subjective evaluation results demonstrate the effectiveness of our proposed methods. Proposed system outperforms conventional cascade approach and the WaveNet based end-to-end approach in terms of both singing quality and singer similarity. Further objective analysis reveals that the model trained with the proposed two-stage training strategy can produce a smoother and sharper formant which leads to higher audio quality.