Background

This is a follow-up to #1100 and PR #1110. The initial fix implemented honoring stt.enabled: false in config, but there are additional improvements needed for a more robust STT (Speech-to-Text) configuration experience.

Current State

PR #1110 added:

• Basic support for stt.enabled: false to skip transcription entirely
• Default behavior unchanged (STT enabled by default)

Proposed Improvements

1. Pluggable STT Providers

Currently, the gateway hardcodes OpenAI Whisper as the STT provider. Users should be able to choose their preferred provider:

Desired providers:

• OpenAI Whisper (current default)
• Deepgram
• Local Whisper (run locally without external API calls)

Proposed config format:

stt:
  enabled: true
  provider: deepgram  # Options: openai, deepgram, local
  openai:
    api_key: ${OPENAI_API_KEY}
    model: whisper-1
  deepgram:
    api_key: ${DEEPGRAM_API_KEY}
    model: nova-2
  local:
    model_path: /path/to/whisper/model
    device: cuda  # or cpu

2. Auto-disable When No API Key is Present

The gateway should gracefully handle missing API keys by:

• Detecting when STT is enabled but no valid API key is configured for the selected provider
• Logging a warning message explaining the situation
• Auto-disabling STT for that session (or permanently until fixed)
• Continuing to operate normally for text-based messages

This prevents the 401 errors described in #1100 when users don't have API keys configured.

3. Runtime Provider Selection (Optional Future)

Consider allowing per-message provider selection via command, e.g.:
/stt use deepgram
/stt use local

Acceptance Criteria

• Config supports stt.provider field with multiple provider options
• Deepgram STT integration implemented
• Local Whisper integration implemented
• Auto-disable logic when API key is missing/absent/invalid
• Clear warning logs when STT is auto-disabled
• Backward compatibility maintained (existing configs still work)
• Tests covering all new provider paths and auto-disable scenarios

Related

• Original issue: STT config ignored: Gateway hardcodes OpenAI Whisper despite stt.enabled: false #1100
• Initial fix PR: fix(gateway): honor stt.enabled: false — skip transcription when STT disabled #1110