Speaches

A self-hosted speech API for transcription, translation, and speech generation, compatible with OpenAI clients and deployable on CPU or GPU.

Speaches is a self-hosted speech server for developers building transcription services and voice applications. Its OpenAI-compatible API lets existing clients and SDKs work with speech models running on hardware you control.

Transcription and translation: Use faster-whisper for speech recognition and translation, with streaming transcription that returns text while processing continues.
Speech generation: Turn text into spoken audio with Kokoro and Piper models.
Audio conversations: Use audio chat completions for spoken summaries, recording sentiment analysis, and asynchronous speech-to-speech exchanges with a model.
Realtime API: Build applications that need ongoing audio interaction through a dedicated Realtime interface.

Speaches runs through Docker or Docker Compose and supports CPU and GPU processing. It loads requested models automatically and unloads them after inactivity, freeing resources between requests. The project is open source under the MIT license.

Translation Inference Servers and Clusters Speech-to-Text and Transcription Text-to-Speech and Audiobooks Docker OpenAI-compatible API Streaming inference