Audio

OpenAI's audio surface has three endpoints. Two are core; one is an extension that's commonly folded into another.

/v1/audio/speech (TTS)

Server takes text, returns audio bytes.

Request:

{
  "model": "<id>",
  "input": "Hello world",
  "voice": "alloy",
  "response_format": "mp3"
}

Response: raw audio bytes. The prober checks:

• Content-Type starts with audio/ (mp3, mpeg, wav, opus, ogg, flac all accepted). application/octet-stream is a WARN — clients usually cope but it's wrong.
• Body length is non-zero.

Common deviations

• response_format ignored. Server always returns MP3 regardless of the requested codec. WARN.
• Voice id required vs. optional. Spec marks voice as required; some servers default to a "voice 0" if omitted. Defaulting is fine; rejecting with 400 is also fine; silently returning empty audio is not.
• Streaming. OpenAI streams audio as the model generates it. Most OSS shims wait for completion and then send the whole file. Catalog treats this as a SKIP for now (no separate stream row); we may add one when streaming TTS becomes common.

/v1/audio/voices (extension — OSS convention)

OpenAI ships a fixed voice enum (alloy, echo, fable, onyx, nova, shimmer) and doesn't expose enumeration. Most OSS TTS servers ship arbitrary voice files (reference-audio-clone TTS, Kokoro, VibeVoice, etc.) and need a way for clients to discover what's available. /v1/audio/voices is the de-facto convention.

Request: GET /v1/audio/voices (no body, no auth-specific params).

Response:

{
  "voices": [
    {"id": "alloy", "name": "Alloy", "language": "en", "sample_rate": 24000},
    {"id": "stevejobs-clone-1", "name": "Steve Jobs (clone)", "language": "en"}
  ]
}

Field	Type	Required	Notes
voices[].id	string	yes	the string to pass as audio.voice on /v1/audio/speech or in [omni] chat
voices[].name	string	no	human-readable display name
voices[].language	string	no	BCP-47 tag (e.g. "en", "en-US")
voices[].sample_rate	integer	no	native sample rate of the voice model
voices[].gender	string	no	implementation-defined enum

A bare-list shape ({"voices": ["alloy", "stevejobs-clone-1"]}) is also seen on minimal servers; clients SHOULD handle either form. The catalog Phase B check requires only that the voices key exists and is non-empty.

Common deviations

• Servers without TTS return 404. The catalog grades this as WARN under the ext kind (capability-gated, not non-compliance).
• Per-voice metadata varies wildly. ElevenLabs returns {voice_id, name, samples, category, ...} (no id field — uses voice_id); most OSS shims pick a flat {id, name, ...}. HT-compat picks id to match /v1/models[i].id.

/v1/audio/transcriptions (STT)

Multipart upload of an audio file, returns transcribed text.

Request fields (multipart):

• file — the audio bytes.
• model — server model id.
• Optional: language, prompt, temperature, response_format (json / text / verbose_json / srt / vtt).

Response (default json):

{ "text": "Hello world" }

Response (verbose_json):

{
  "text": "Hello world",
  "language": "en",
  "duration": 2.4,
  "segments": [...]
}

The prober probes the default json format with one second of silent WAV — enough to exercise the upload path without burning model inference budget on a real signal.

Common deviations

• text field but body wrapped ({"transcription": {"text": "..."}}). Hard FAIL.
• No language even on verbose_json. WARN — some Whisper variants return verbose_json shape but skip language detection when given silent audio. The prober treats it leniently.
• 422 on a 1-second probe. A few servers reject sub-N-second uploads as "too short to transcribe". Catalog accepts a 422 here as a documented deviation (PASS-with-warning) since the route clearly exists.

/v1/audio/translations (STT to English)

ext. OpenAI keeps it; many OSS servers fold its functionality into /v1/audio/transcriptions with a task: translate parameter. A 404 here is SKIP, not FAIL.