Speaches
API toolsA compatible API for speech models.
Speaches packages open speech models behind endpoints shaped for existing OpenAI clients. A self-hosted deployment can transcribe or translate recordings, stream partial transcripts, synthesize speech, run real-time voice sessions, and load or unload models dynamically on CPU or GPU infrastructure.
- Transcribe and translate audio with streaming OpenAI-compatible endpoints
- Generate speech with Kokoro and Piper models
- Run real-time voice, diarization, model management, and voice activity detection
What people use it for
Replace a hosted speech endpoint
Point compatible clients at a self-hosted Speaches base URL to transcribe, translate, or synthesize audio while keeping model execution on infrastructure selected by the operator.
Stream live transcription
Receive incremental transcription results over server-sent events or use the Realtime API in transcription-only mode for live subtitles, meeting notes, voice capture, and accessibility workflows.
Prototype a real-time voice assistant
Combine speech recognition, an external conversation model, and speech synthesis over the OpenAI-compatible WebSocket interface to test interactive voice applications with configurable local audio models.