Voice: speech ↔ text

Local transcription (STT) and natural speech (TTS) with mlx-audio.

Bedrova does both directions of voice locally, powered by mlx-audio.

Screenshot: the voice view with a live transcript — added at launch.

Speech-to-text

Whisper, Distil-Whisper, Parakeet, and other STT models. Transcribe from the microphone or a file, with timestamps, voice-activity detection (VAD), diarization, and streaming transcription. Audio stays on your Mac by default.

Text-to-speech

Natural voices via Kokoro, Qwen3-TTS, and others, with voice selection and streaming synthesis. Speech-to-speech is supported where the model allows.

API

OpenAI-compatible POST /v1/audio/transcriptions and POST /v1/audio/speech. List available voices via GET /api/v1/audio/voices. Like every model type, voice models load just-in-time and share the memory budget.