PaddleSpeech

An open-source speech AI toolkit for local or self-hosted use, with streaming recognition, text-to-speech and Chinese pronunciation processing.

Screenshot of PaddleSpeech website

PaddleSpeech is a speech and audio toolkit for developers and researchers building applications on their own machines or servers. Built on PaddlePaddle, it supports pretrained models alongside training and evaluation workflows.

  • Speech recognition: Transcribe audio with models including Whisper large v3 and turbo, Conformer and DeepSpeech2, with streaming recognition and punctuation restoration.
  • Speech synthesis: Generate speech with Tacotron2, FastSpeech2 or ERNIE-SAT and vocoders such as HiFiGAN, including support for streaming text-to-speech.
  • Chinese text processing: Normalize written text and resolve pronunciation differences caused by characters with multiple readings or tone changes in context.
  • Audio analysis and translation: Verify speakers, detect keywords, classify audio and translate English speech into Chinese.
  • Specialized workflows: Use code-switching recognition, SRT subtitle generation examples and singing voice synthesis examples based on DiffSinger.

PaddleSpeech is open source under Apache 2.0. It runs on Linux, Windows and macOS, with Linux recommended, and supports CPU execution. Server interfaces support self-hosted speech applications; its Python modules also let researchers train and test models within the toolkit.

Similar to PaddleSpeech