Parallel WaveNet: Fast High-Fidelity Speech Synthesis
read the original abstract
The recently-developed WaveNet architecture is the current state of the art in realistic speech synthesis, consistently rated as more natural sounding for many different languages than any previous system. However, because WaveNet relies on sequential generation of one audio sample at a time, it is poorly suited to today's massively parallel computers, and therefore hard to deploy in a real-time production setting. This paper introduces Probability Density Distillation, a new method for training a parallel feed-forward network from a trained WaveNet with no significant difference in quality. The resulting system is capable of generating high-fidelity speech samples at more than 20 times faster than real-time, and is deployed online by Google Assistant, including serving multiple English and Japanese voices.
This paper has not been read by Pith yet.
Forward citations
Cited by 8 Pith papers
-
VideoGPT: Video Generation using VQ-VAE and Transformers
VideoGPT generates competitive natural videos by learning discrete latents with VQ-VAE and modeling them autoregressively with a transformer.
-
Generalization of Spectrum Differential based Direct Waveform Modification for Voice Conversion
Residual-domain F0 transformation generalizes spectrum-differential direct waveform modification to arbitrary spectral conversion models in voice conversion.
-
EPNAS: Efficient Progressive Neural Architecture Search
EPNAS uses a progressive search policy with REINFORCE performance prediction to search neural architectures in parallel, supporting multiple resource constraints and outperforming ENAS and PNAS on CIFAR-10 and ImageNe...
-
Phoneme-Based Contextualization for Cross-Lingual Speech Recognition in End-to-End Models
An E2E ASR model with mixed wordpieces and phonemes improves foreign proper noun recognition via phoneme-level contextual biasing, showing 16% gain over grapheme-only and 8% over wordpiece-only baselines.
-
Hierarchical Sequence to Sequence Voice Conversion with Limited Data
Hierarchical seq2seq model for parallel voice conversion pretrained as autoencoder on single-speaker data then adapted to limited multispeaker data, using mel spectrograms converted via wavenet vocoder.
-
Autoencoding sensory substitution
Deep recurrent autoencoders convert images to shortened audio signals that incorporate hearing models, enabling above-chance hand posture discrimination and object reaching after a few hours of training instead of months.
-
Improving Performance of End-to-End ASR on Numeric Sequences
TTS-generated numeric training data plus a compact neural denormalizer improve E2E ASR word error rates on numeric sequences by up to a factor of 8 for the longest cases.
-
Sequence Generation: From Both Sides to the Middle
SBSG model generates sequences bidirectionally from ends to middle via interactive attention, claiming faster decoding and better quality than autoregressive Transformer on NMT and summarization tasks.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.