Training Neural Speech Recognition Systems with Synthetic Speech Augmentation
arXiv:1811.00707
Abstract
Building an accurate automatic speech recognition (ASR) system requires a large dataset that contains many hours of labeled speech samples produced by a diverse set of speakers. The lack of such open free datasets is one of the main issues preventing advancements in ASR research. To address this problem, we propose to augment a natural speech dataset with synthetic speech. We train very large end-to-end neural speech recognition models using the LibriSpeech dataset augmented with synthetic speech. These new models achieve state of the art Word Error Rate (WER) for character-level based models without an external language model.
arXiv commentPre-print. Work in progress, 5 pages, 1 figure
References in corpus (5)
- WaveNet: A Generative Model for Raw Audio
- Deep Speech 2: End-to-End Speech Recognition in English and Mandarin
- Style Tokens: Unsupervised Style Modeling, Control and Transfer in End-to-End Speech Synthesis
- Improved training of end-to-end attention models for speech recognition
- Wav2Letter: an End-to-End ConvNet-based Speech Recognition System
Cited by in corpus (8)
- Self-Supervised Speech Representation Learning: A Review
- Augmented Datasheets for Speech Datasets and Ethical Decision-Making
- End-to-End Spoken Language Understanding: Performance analyses of a voice command task in a low resource setting
- Text-only domain adaptation for end-to-end ASR using integrated text-to-mel-spectrogram generator
- Training Keyword Spotters with Limited and Synthesized Speech Data
- Measuring the Effectiveness of Voice Conversion on Speaker Identification and Automatic Speech Recognition Systems
- Injecting Text in Self-Supervised Speech Pretraining
- Using Synthetic Audio to Improve The Recognition of Out-Of-Vocabulary Words in End-To-End ASR Systems