FastSpeech: Parallel Mel-Spectrogram Generation for Text-to-Speech

FastSpeech is a neural text-to-speech model developed by Yi Ren of Zhejiang University. The model uses a feed-forward Transformer network to generate mel-spectrograms in parallel, addressing speed and robustness issues in autoregressive TTS. Experiments on the LJSpeech dataset show the model speeds up mel-spectrogram generation by 270x and end-to-end synthesis by 38x compared to an autoregressive Transformer baseline.

Use Cases

• Training fast TTS models based on the described parallel feed-forward Transformer architecture.
• Benchmarking speech synthesis speed and quality against autoregressive models like Tacotron 2.
• Controlling voice speed and prosody in synthesized speech based on the model's described length regulator mechanism.
• Investigating robustness against word skipping and repeating in TTS outputs as mentioned in the description.

Strengths

• Demonstrates a 270x speedup in mel-spectrogram generation and a 38x speedup in end-to-end synthesis compared to an autoregressive Transformer TTS model.
• Reported to nearly eliminate word skipping and repeating in hard cases and allows smooth adjustment of voice speed.
• Based on the LJSpeech dataset, a standard benchmark for TTS research.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and dataset size are unknown, which may limit suitability assessment.
• Last update date is unknown; freshness unverified.

Provenance

Source

Zhejiang University

Collection Method

Neural network model trained on the LJSpeech dataset.