Speech & Transcription
Two endpoints, both shaped like OpenAI's, so an existing client works by changing the base URL: /audio/speech turns text into audio, and /audio/transcriptions turns audio into text.
Text to speech
POST /api/v1/audio/speech responds with the audio itself, not JSON, so you can pipe it straight into a file or a player.
curl https://mume.ai/api/v1/audio/speech \
-H "Authorization: Bearer $MUME_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "microsoft/mai-voice-2",
"voice": "en-US-Harper",
"input": "The build finished, and every test passed."
}' \
--output reply.mp3Parameters
| Field | Required | Notes |
|---|---|---|
model | Yes | Any model whose surface is speech. See the catalogue. |
input | Yes | The text to read. Must be non-empty. |
voice | Yes | A voice the model actually has. This is required here even though some upstreams treat it as optional — see the note below. |
response_format | No | Defaults to mp3. The response's Content-Type tells you what you actually got. |
store | No | Defaults to on. Set false to skip saving the clip to your media library. |
stream | No | Defaults to off. Set true to receive the audio as it is synthesised — see below. |
Streaming
Most speech models generate audio incrementally, and stream: true passes that through instead of waiting for the clip to finish. It is the same response — same headers, same bytes, same price — the first of them just arrives much sooner. On deepgram/aura-2, a paragraph that takes 32 seconds to synthesise starts arriving after 2.
curl https://mume.ai/api/v1/audio/speech \
-H "Authorization: Bearer $MUME_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "deepgram/aura-2",
"voice": "aura-2-thalia-en",
"input": "Streaming starts playing before synthesis has finished.",
"stream": true
}' \
--no-buffer --output - | ffplay -nodisp -Not every model streams: some synthesise the whole clip upstream and send it in one piece. Asking for a stream from one of those is harmless — you get the same audio, all at once — so there is nothing to look up and no reason to special-case it.
A failure after the first byte closes the connection. The status line has already been sent by then, so there is no error code left to give you. A truncated transfer is the signal; treat a clip that ends early as failed rather than short. Nothing else changes — X-Media-Url is still sent up front, and still points at the complete clip once it lands.
A voice belongs to a model
Voice names are not portable — en-US-Harper exists on one model and not on another — so the gateway requires voice rather than choosing for you. Omitting it upstream produces a schema error that reads like a problem with your audio settings, which is a confusing way to learn that you forgot a field.
Match the voice to the language of the text. Nothing stops a Spanish voice reading French, and the result is the most common way this goes wrong.
Response headers
A speech response is the audio, so there is nowhere in the body to report a cost or an id. Both come back as headers — and because they are headers, they are already there when a streamed clip starts arriving:
| Header | Meaning |
|---|---|
X-Generation-Id | Identifies the clip. Always present. |
X-Media-Url | The hosted copy in your media library. Absent when you passed store: false. |
Speech to text
POST /api/v1/audio/transcriptions accepts either the OpenAI multipart shape or a JSON body carrying a URL. Both return the same thing.
Multipart, with a file
curl https://mume.ai/api/v1/audio/transcriptions \
-H "Authorization: Bearer $MUME_API_KEY" \
-F model="openai/whisper-1" \
-F file=@interview.mp3JSON, with a URL
The JSON form exists because callers often already hold a URL — a file in their own storage, or something a user uploaded. Downloading it only to upload it again through us is pure waste.
curl https://mume.ai/api/v1/audio/transcriptions \
-H "Authorization: Bearer $MUME_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "openai/whisper-1",
"file_url": "https://example.com/interview.mp3"
}'Response
{
"text": "So the thing about the migration is that we never turned the old path off…",
"usage": {
"seconds": 184.2,
"cost": 0.0011
}
}Limits
- 25 MB per file. Check the size before you send — a rejected upload costs you the round trip.
- Container formats the model accepts go straight through, including
mp4andwebm. The gateway does not transcode, so it cannot rescue a format the model refuses.
Errors
Both endpoints use the shared error shape — see Error Handling. The two you will meet first are invalid_request for a missing model, input or voice, and 402 when the account is out of credit.
Related
- Images & Video — the other two media surfaces, which bill the same way.
- Voice-over Translator — an agent built on both of these endpoints, if you would rather not wire them together yourself.