arXiv · 2021 · Preprint
Ravanelli et al. · → Paper · Demo: ? · Code: ✓
SpeechBrain is an open-source, all-in-one PyTorch-based speech toolkit providing unified training recipes, pretrained models, and evaluation infrastructure for ASR, speaker recognition, speech enhancement, and speech separation.
Citation Stub
This paper is not a speech generation paper but is cited by the corpus. See Context in Speech Generation below for why it is relevant.
Context in Speech Generation
SpeechBrain provides a modular, PyTorch-native framework for training and evaluating speech processing models across multiple tasks including ASR, speaker verification, speaker diarization, speech enhancement, and speech separation. Its pretrained ASR models (trained on LibriSpeech, CommonVoice, and TIMIT) and speaker embedding models (including ECAPA-TDNN) can serve as off-the-shelf evaluators, enabling TTS and voice conversion pipelines to measure intelligibility via WER and speaker similarity via EER without training their own evaluation models from scratch. The toolkit’s design emphasises a single unified codebase, with task-specific recipes sharing common data loading, batching, and training infrastructure, making it straightforward to integrate speech evaluation components into research pipelines. At release, TTS support was listed as forthcoming; the toolkit’s primary value to the speech generation community lies in its ASR and speaker embedding components, which provide a reproducible and openly licensed evaluation backbone.
Wiki Connections
- evaluation-metrics — SpeechBrain’s pretrained ASR (WER) and speaker embedding (EER/SPK-SIM) models are used as off-the-shelf evaluators in TTS and voice conversion pipelines