Local Audio Setup

May 16, 2026 ยท View on GitHub

OwnPilot can run speech-to-text and text-to-speech without paid APIs.

Components

  • STT: whisper.cpp running an OpenAI-compatible server.
  • TTS: piper CLI with a local .onnx voice model.

Config Center

Create or update audio_service:

provider_type = local
base_url = http://127.0.0.1:2022
local_tts_command = piper
local_tts_model = D:\models\piper\tr_TR-voice.onnx

api_key is not required for local mode.

Run Local STT

Run whisper.cpp in OpenAI-compatible server mode and point base_url at it. The server must expose:

POST /v1/audio/transcriptions

Run Local TTS

Install Piper and download a voice model. OwnPilot calls:

piper --model <local_tts_model> --output_file <temp.wav>

Piper output is WAV. Telegram voice replies try to convert WAV to OGG/Opus with ffmpeg so they can be sent as native voice notes. If ffmpeg is unavailable, OwnPilot falls back to sending the generated audio file.

Telegram

For voice input only, set audio_service.provider_type = local.

For voice replies, also set Telegram config:

voice_reply_mode = voice_messages

Use always only if every assistant response should be synthesized.

Diagnostics

After saving Config Center settings, check readiness:

GET /api/v1/voice/diagnostics

For local mode this verifies:

  • base_url responds for the local Whisper server.
  • local_tts_model points to an existing Piper .onnx file.
  • local_tts_command can be executed.
  • ffmpeg is available for Telegram native voice replies.

ffmpeg is optional. If it is missing, Telegram TTS replies still work as audio file attachments instead of native voice notes.