Add speech-to-text, text-to-speech, and ElevenLabs provider

[patrickdet]

Summary

Adds TTS and STT to req_llm, plus an ElevenLabs provider.

Speech (ReqLLM.Speech)

Text-to-speech through the standard prepare_request(:speech, ...) pipeline. Any provider that implements the operation works.

{:ok, result} = ReqLLM.speak("openai:tts-1", "Hello world", voice: "alloy")
File.write!("hello.mp3", result.audio)

Options: voice selection, speed, output format (mp3/wav/opus/flac/aac/pcm), language hints, provider-specific stuff like OpenAI's instructions for gpt-4o-mini-tts.

Transcription (ReqLLM.Transcription)

Speech-to-text. Takes file paths, raw binary, or base64 audio. Returns text with optional segment timing.

{:ok, result} = ReqLLM.transcribe("groq:whisper-large-v3-turbo", "recording.mp3")
result.text #=> "Hello world"
result.segments #=> [%{text: "Hello world", start_second: 0.0, end_second: 1.2}]

ElevenLabs provider

Speech-only. Their API is pretty different from OpenAI's /audio/speech:

• Voice ID in the URL path (/v1/text-to-speech/{voiceId})
• xi-api-key header instead of Bearer auth
• Output format as a query param, not body field
• text/model_id instead of input/model

voice_settings (stability, similarity_boost, style, speed) go through provider_options. Auto-discovered at startup.

{:ok, result} = ReqLLM.speak(
  %{id: "eleven_multilingual_v2", provider: :elevenlabs},
  "Hello!",
  provider_options: [stability: 0.5, similarity_boost: 0.8]
)

Integration tests

Tagged :integration, excluded by default. Tested against real APIs:

• ElevenLabs TTS: default voice, voice_settings, language codes
• OpenAI TTS: tts-1, wav output
• Groq STT: generate-then-transcribe pattern (OpenAI TTS makes audio, Groq whisper transcribes it) so we don't commit binary fixtures

ELEVENLABS_API_KEY=... OPENAI_API_KEY=... GROQ_API_KEY=... \
  mix test --include integration test/req_llm/integration/

Test plan

• ElevenLabs unit tests pass (18 tests)
• Full suite passes (2370 tests, 0 failures)
• Integration tests pass against real ElevenLabs, OpenAI, and Groq APIs (8/8)

[mikehostetler]

Blocking on agentjido/llm_db#152 for first-class ElevenLabs support. The provider module can work with explicit map specs, but string model specs like elevenlabs:eleven_multilingual_v2 still depend on llm_db provider/model metadata. Until that lands, ElevenLabs remains a custom-provider path rather than a normal built-in provider.

I'm actively working on this issue - so it won't be a blocker for long

[mikehostetler]

Status update as of March 10, 2026:

• Updated the local PR branch to use llm_db from GitHub main so we can validate against the merged ElevenLabs support before the next Hex release.
• Added the ElevenLabs provider to this repo's local llm_db allow-filter so elevenlabs:* models resolve correctly in test/runtime config.
• Extended the PR branch's ElevenLabs provider to support speech-to-text via POST /v1/speech-to-text in addition to the existing TTS path.
• Updated transcription parsing to handle ElevenLabs response fields like language_code, words[].text, and multichannel transcript payloads.
• Added unit/integration coverage for the ElevenLabs STT path.

Verification:

• mix test test/req_llm/speech_test.exs test/req_llm_test.exs test/req_llm/providers/elevenlabs_test.exs test/req_llm/transcription_test.exs passes locally.
• ReqLLM.model("elevenlabs:eleven_multilingual_v2") now resolves correctly against the GitHub llm_db branch.
• Live ElevenLabs verification is still blocked in this environment by an upstream HTTP 401 / invalid_api_key response from ElevenLabs, not by the request/response wiring in req_llm.

We should keep this PR in draft until llm_db is released, since landing this against an unreleased Hex dependency would create a merge timing problem for main.

[mikehostetler]

Status update as of March 10, 2026:

• Updated this PR to use released Hex llm_db 2026.3.1 instead of the temporary GitHub main override.
• Rebases cleanly on current main.
• Kept the ElevenLabs speech-to-text, transcription parsing, model resolution, and coverage changes on the branch.

Verification:

• mix test passes locally.
• mix format --check-formatted passes locally.
• mix test test/req_llm/speech_test.exs test/req_llm_test.exs test/req_llm/providers/elevenlabs_test.exs test/req_llm/transcription_test.exs passes locally.

Remaining limitation in this environment:

• Live ElevenLabs STT integration was not rerun here because ELEVENLABS_API_KEY is not available locally.

This should now be ready for final review.