Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Glow-TTS: A Generative Flow for Text-to-Speech via Monotonic Alignment Search

Jaehyeon Kim, Sungwon Kim, Jungil Kong, Sungroh Yoon

TL;DR

Glow-TTS proposes a flow-based parallel TTS model that learns monotonic text-to-speech alignments internally by combining a probabilistic flow decoder with a dynamic-programming monotonic alignment search. The training objective maximizes the data likelihood $\log P_{X}(x|c; \theta, A)$ by iteratively finding the best alignment $A^*$ via MAS and updating model parameters, enabling fast, parallel generation and robustness to long utterances. The approach achieves substantial synthesis speedups over autoregressive baselines, while maintaining comparable speech quality, and provides controllability over prosody and speaking rate; it also extends naturally to multi-speaker settings and voice conversion. This combination of hard monotonic alignment, exact likelihood training, and efficient parallel decoding offers a practical and scalable alternative for high-quality, low-latency TTS systems.

Abstract

Recently, text-to-speech (TTS) models such as FastSpeech and ParaNet have been proposed to generate mel-spectrograms from text in parallel. Despite the advantage, the parallel TTS models cannot be trained without guidance from autoregressive TTS models as their external aligners. In this work, we propose Glow-TTS, a flow-based generative model for parallel TTS that does not require any external aligner. By combining the properties of flows and dynamic programming, the proposed model searches for the most probable monotonic alignment between text and the latent representation of speech on its own. We demonstrate that enforcing hard monotonic alignments enables robust TTS, which generalizes to long utterances, and employing generative flows enables fast, diverse, and controllable speech synthesis. Glow-TTS obtains an order-of-magnitude speed-up over the autoregressive model, Tacotron 2, at synthesis with comparable speech quality. We further show that our model can be easily extended to a multi-speaker setting.

Glow-TTS: A Generative Flow for Text-to-Speech via Monotonic Alignment Search

TL;DR

Glow-TTS proposes a flow-based parallel TTS model that learns monotonic text-to-speech alignments internally by combining a probabilistic flow decoder with a dynamic-programming monotonic alignment search. The training objective maximizes the data likelihood by iteratively finding the best alignment via MAS and updating model parameters, enabling fast, parallel generation and robustness to long utterances. The approach achieves substantial synthesis speedups over autoregressive baselines, while maintaining comparable speech quality, and provides controllability over prosody and speaking rate; it also extends naturally to multi-speaker settings and voice conversion. This combination of hard monotonic alignment, exact likelihood training, and efficient parallel decoding offers a practical and scalable alternative for high-quality, low-latency TTS systems.

Abstract

Recently, text-to-speech (TTS) models such as FastSpeech and ParaNet have been proposed to generate mel-spectrograms from text in parallel. Despite the advantage, the parallel TTS models cannot be trained without guidance from autoregressive TTS models as their external aligners. In this work, we propose Glow-TTS, a flow-based generative model for parallel TTS that does not require any external aligner. By combining the properties of flows and dynamic programming, the proposed model searches for the most probable monotonic alignment between text and the latent representation of speech on its own. We demonstrate that enforcing hard monotonic alignments enables robust TTS, which generalizes to long utterances, and employing generative flows enables fast, diverse, and controllable speech synthesis. Glow-TTS obtains an order-of-magnitude speed-up over the autoregressive model, Tacotron 2, at synthesis with comparable speech quality. We further show that our model can be easily extended to a multi-speaker setting.

Paper Structure

This paper contains 15 sections, 10 equations, 8 figures, 5 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Glow-TTS
  4. Training and Inference Procedures
  5. Monotonic Alignment Search
  6. Model Architecture
  7. Decoder.
  8. Encoder and Duration Predictor.
  9. Experiments
  10. Results
  11. Audio Quality
  12. Sampling Speed and Robustness
  13. Diversity and Controllability
  14. Multi-Speaker TTS
  15. Conclusion

Figures (8)

  • Figure 1: Training and inference procedures of Glow-TTS.
  • Figure 2: Illustrations of the monotonic alignment search.
  • Figure 3: Comparison of inference time and length robustness.
  • Figure 4: The fundamental frequency (F0) contours of synthesized speech samples from Glow-TTS trained on the LJ dataset.
  • Figure 5: Mel-spectrograms of the generated speech samples with different speaking rates.
  • ...and 3 more figures