Context
ccbot's voice-message transcription (src/ccbot/transcribe.py) hardcodes the model:
data={"model": "gpt-4o-transcribe"},OPENAI_BASE_URL already makes the endpoint configurable, so ccbot can target any OpenAI-compatible /audio/transcriptions server — but the hardcoded model name blocks backends that serve different model names (e.g. a self-hosted Whisper server).
Use case
Private / offline transcription against a local OpenAI-compatible backend. I run Speaches on my LAN exposing Whisper (large-v3-turbo, via faster-whisper) at http://<host>:9000/v1/audio/transcriptions. It's a drop-in OpenAI Whisper-compatible endpoint, but ccbot always requests gpt-4o-transcribe, which that server doesn't serve.
Proposal
Add a CCBOT_TRANSCRIBE_MODEL env var (default gpt-4o-transcribe), mirroring the existing OPENAI_BASE_URL / OPENAI_API_KEY overrides. Backwards compatible: unset preserves current OpenAI behaviour.
Verified end-to-end: a Telegram voice message → Speaches → correctly transcribed text forwarded to the Claude session.
PR incoming.
Context
ccbot's voice-message transcription (
src/ccbot/transcribe.py) hardcodes the model:OPENAI_BASE_URLalready makes the endpoint configurable, so ccbot can target any OpenAI-compatible/audio/transcriptionsserver — but the hardcoded model name blocks backends that serve different model names (e.g. a self-hosted Whisper server).Use case
Private / offline transcription against a local OpenAI-compatible backend. I run Speaches on my LAN exposing Whisper (
large-v3-turbo, via faster-whisper) athttp://<host>:9000/v1/audio/transcriptions. It's a drop-in OpenAI Whisper-compatible endpoint, but ccbot always requestsgpt-4o-transcribe, which that server doesn't serve.Proposal
Add a
CCBOT_TRANSCRIBE_MODELenv var (defaultgpt-4o-transcribe), mirroring the existingOPENAI_BASE_URL/OPENAI_API_KEYoverrides. Backwards compatible: unset preserves current OpenAI behaviour.Verified end-to-end: a Telegram voice message → Speaches → correctly transcribed text forwarded to the Claude session.
PR incoming.