[Paper Review] AdaDurIAN: Few-shot Adaptation for Neural Text-to-Speech with DurIAN

Zewang Zhang, Qiao Tian+3 more|arXiv (Cornell University)|May 12, 2020

Natural Language Processing TechniquesComputer Science27 references22 citations

TL;DR

AdaDurIAN proposes a few-shot adaptation method for neural text-to-speech using an improved DurIAN-based average model, enabling high-quality, natural, and speaker-similar speech synthesis with only minutes of monolingual data. By fine-tuning only the speaker embedding and decoder while keeping the content encoder fixed, it achieves superior naturalness (MOS 4.21 with 3 minutes of data) and robust cross-lingual performance compared to baseline Tacotron-like models.

ABSTRACT

This paper investigates how to leverage a DurIAN-based average model to enable a new speaker to have both accurate pronunciation and fluent cross-lingual speaking with very limited monolingual data. A weakness of the recently proposed end-to-end text-to-speech (TTS) systems is that robust alignment is hard to achieve, which hinders it to scale well with very limited data. To cope with this issue, we introduce AdaDurIAN by training an improved DurIAN-based average model and leverage it to few-shot learning with the shared speaker-independent content encoder across different speakers. Several few-shot learning tasks in our experiments show AdaDurIAN can outperform the baseline end-to-end system by a large margin. Subjective evaluations also show that AdaDurIAN yields higher mean opinion score (MOS) of naturalness and more preferences of speaker similarity. In addition, we also apply AdaDurIAN to emotion transfer tasks and demonstrate its promising performance.

Motivation & Objective

• To address the challenge of training high-quality neural TTS systems with extremely limited monolingual data for new speakers.
• To improve robustness and naturalness in few-shot speaker adaptation by leveraging duration-aware attention and a shared speaker-independent content encoder.
• To enable fluent cross-lingual speech synthesis in a new speaker's voice using only minutes of data, overcoming instability in attention-based end-to-end TTS models.
• To evaluate the model’s performance in both speaker adaptation and few-shot emotion transfer tasks.

Proposed method

• Proposes AdaDurIAN, a few-shot adaptation framework based on DurIAN, using a shared speaker-independent content encoder for phoneme and tone/stress sequences.
• Fixes the content encoder and only fine-tunes the speaker embedding and decoder head, reducing training instability and pronunciation errors.
• Introduces a time-delayed LSTM post-net for streaming inference, replacing the global CBHG module to improve mel-spectrogram smoothness.
• Employs windowed content-based attention with duration supervision to enhance alignment robustness and reduce attention collapse.
• Trains a multi-speaker average model and adapts it via fine-tuning with only 1–20 minutes of monolingual data per speaker.
• Applies the same adaptation strategy to emotion transfer tasks, transferring emotions from a neutral base model with minimal data.

Experimental results

Research questions

• RQ1 Can a single average TTS model be effectively adapted to new speakers using only minutes of monolingual data while maintaining high naturalness and speaker similarity?
• RQ2 How does fine-tuning only the speaker embedding and decoder, while keeping the content encoder fixed, affect pronunciation accuracy and alignment robustness in few-shot settings?
• RQ3 To what extent can AdaDurIAN generalize to cross-lingual speech synthesis when the target speaker has no prior data in the target language?
• RQ4 Can AdaDurIAN be successfully applied to few-shot emotion transfer tasks, and how does its performance compare across different emotion styles?
• RQ5 How does AdaDurIAN’s performance compare to standard Tacotron-like models in terms of MOS, speaker similarity, and cross-lingual fluency?

Key findings

• With only 3 minutes of training data, AdaDurIAN achieved a MOS of 4.21 for a native Chinese speaker, only 0.2 below the ground-truth recording MOS.
• For a native English speaker with no Chinese data, AdaDurIAN achieved a MOS of 3.90 on Chinese sentences with just 1 minute of data, showing strong cross-lingual generalization.
• AdaDurIAN outperformed the SMA baseline in speaker similarity preference tests, especially for the Chinese speaker, with a significant margin on English sentences.
• The model achieved a 64% mean emotion classification accuracy in few-shot F2F emotion transfer and 52% in F2M transfer, demonstrating promising performance in emotion transfer with minimal data.
• Subjective evaluations confirmed that AdaDurIAN yields higher naturalness and greater preference for speaker similarity than the baseline Tacotron-like model.
• The method significantly reduces pronunciation errors and attention collapse by fixing the content encoder and only fine-tuning the speaker embedding and decoder.