Favicon of SpeechBrain

SpeechBrain

An open-source speech AI toolkit supports transcription, audio cleanup and custom model training with PyTorch and Hugging Face under Apache 2.0.

Screenshot of SpeechBrain website

SpeechBrain is a toolkit for researchers and developers building speech assistants, transcription systems and conversational AI. It provides pretrained models and training recipes for adapting speech, audio and text models to your own data.

  • Speech recognition and generation: Build transcription, speaker recognition, text-to-speech and speech translation systems, including speech-to-speech translation.
  • Audio cleanup and analysis: Reduce noise, separate overlapping voices, detect sound events and process recordings from multiple microphones.
  • Custom model training: Train from scratch or fine-tune Whisper, Wav2Vec2, WavLM and Hubert, with control over models, losses and evaluation.
  • Text and dialogue: Train language models and response generators using models such as GPT2 and Llama2, and combine them with speech processing pipelines.

SpeechBrain runs as a Python and PyTorch toolkit and is open source under Apache 2.0. Hugging Face hosts pretrained models for tasks including transcription, speaker verification and speech enhancement. Training supports single or multiple GPUs, distributed jobs and mixed precision. Dataset recipes, published training logs and checkpoints help researchers compare experiments against reference results.

Similar to SpeechBrain