Add AddisAI STT and TTS plugin

[ZelalemGizachew]

What this adds

I'm adding an AddisAI plugin for LiveKit Agents with:

• batch STT through addis-whisper
• non-streaming TTS through Addis Voices 2
• explicit Amharic (am) and Afaan Oromo (om) language support
• dynamic voice IDs, with am-hamen as the documented Amharic default
• pcm_16000, wav_44100, and mp3_44100 output support

AddisAI is building language-native speech models for Amharic and Afaan Oromo, and I wanted those services to be usable directly in a LiveKit voice pipeline.

Implementation notes

The current AddisAI Python SDK is synchronous, so this plugin uses LiveKit's shared asynchronous HTTP session instead of blocking the agent event loop.

The capabilities match what the APIs provide today:

• STT is marked non-streaming and returns final transcripts. LiveKit can adapt it with a VAD.
• TTS is marked non-streaming. LiveKit can adapt it by synthesizing sentence chunks.
• TTS keeps one provider client_request_id across retry attempts so an uncertain response does not create or bill a second clip.
• Every retry attempt uses a fresh LiveKit output request ID, while the stable provider clip ID is retained separately for trace correlation. This lets downstream consumers discard partial audio from a failed attempt safely.
• STT maps the provider request ID, confidence, and usage metadata.
• The STT multipart field headers match the official SDK wire format. The production endpoint ignores request_data when that part is labeled text/plain.
• Provider HTTP failures remain LiveKit APIStatusError values so retryability and diagnostics are preserved.

I intentionally kept AddisAI text/multimodal generation and the realtime audio WebSocket out of this PR. They are promising follow-ups, but their current public contracts do not yet expose all of the streaming, tool, transcript, session, and interruption behavior needed for complete LiveKit LLM or Realtime Model implementations.

Validation

• 13 hermetic unit tests covering STT multipart requests, response mapping, language validation, HTTP errors, TTS clip download, inline audio, retry idempotency, and fresh output IDs across retries
• Ruff formatting and linting
• Mypy type checking
• Wheel and source-distribution builds
• Live Amharic round trip: am-hamen TTS to STT, 0.919 transcription confidence
• Live Afaan Oromo round trip: om-ayyaantuu TTS to STT, 0.949 transcription confidence
• End-to-end LiveKit Cloud test project using the AddisAI STT and TTS providers

The credential-backed smoke tests exercised the plugin itself against the production AddisAI TTS and STT endpoints. Both language paths completed successfully, and I have also tested the providers in a LiveKit Cloud agent, so I'm marking this ready for review.

Node.js companion

The matching Node.js package is now proposed separately in livekit/agents-js#2160. It follows the agents-js package architecture as @livekit/agents-plugin-addisai and includes the same STT/TTS language support, retry behavior, tests, TypeDoc/API report, README, and a Gemini example.

Documentation and provider listing

I would also like AddisAI to be discoverable alongside the other model providers after these plugins are merged and released. Could you please:

• add AddisAI (Python and Node.js) to the STT provider table at /agents/models/stt/
• add AddisAI (Python and Node.js) to the TTS provider table at /agents/models/tts/
• publish an AddisAI STT guide at /agents/models/stt/addisai/ with Python and Node.js tabs
• publish an AddisAI TTS guide at /agents/models/tts/addisai/ with Python and Node.js tabs

The Python package README and Node package README/API report include docs-ready installation, authentication, usage, parameters, language support, and resource links for both guides. I'm happy to send a separate docs PR too if the current documentation source or workflow is available to external contributors.

[CLAassistant]

All committers have signed the CLA.

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.

[ZelalemGizachew]

@codex Please resolve the current merge conflict against the latest main in uv.lock on this PR branch. Merge/rebase the latest main, preserve the AddisAI workspace member and package metadata, regenerate the lockfile with uv lock rather than manually editing generated dependency entries, run the relevant formatting/type/unit checks, and commit the resolution to this PR branch.

[chatgpt-codex-connector]

You have reached your Codex usage limits. You can see your limits in the Codex usage dashboard.