Skip to main content
POST
Transcribe an audio file through the attested TEE gateway using multipart/form-data. Responses are buffered JSON with an X-Receipt-Id header for verification.

Request body

string
required
Transcription model id, for example openai/whisper-large-v3; see the models catalog (GET /v1/models) for available ids. Transcription models have output_modalities: ["transcription"] in the catalog.
file
required
Exactly one nonempty file, up to 25 MiB (26,214,400 bytes). The filename must include one of these extensions: flac, mp3, mp4, mpeg, mpga, m4a, ogg, wav, or webm. qwen/qwen3-asr-1.7b and qwen/qwen3-asr-0.6b accept up to 600 seconds of audio per request and return 400 invalid_request_error for longer files.
string
Language code of 2–3 lowercase ASCII letters, for example en. The Qwen3-ASR models detect the language themselves and may ignore this hint.
string
Optional context for transcription, up to 4,096 Unicode characters.
number
Sampling temperature from 0 through 1.
string
json (default) is supported by every model. verbose_json is supported only by openai/whisper-large-v3; mistralai/voxtral-small-24b-2507, qwen/qwen3-asr-1.7b and qwen/qwen3-asr-0.6b return 400 invalid_request_error for verbose_json. Text, subtitle, and other response formats are not supported.
string[]
word and/or segment, once each, only with openai/whisper-large-v3 and response_format=verbose_json. Repeat the field to request both. timestamp_granularities is also accepted; do not mix spellings. Word-level timestamps are not returned; segments may be absent.
boolean
Omit this field or send the literal form value false. Streaming is not supported; true and all other values return 400.
string
Accepted for SDK compatibility but removed before forwarding upstream.
Duplicate scalar fields, malformed multipart, and invalid UTF-8 text return 400. Other fields, including provider, chunking_strategy, include[], and speaker fields, return 400 unknown_parameter. ACI E2EE is not supported on this endpoint. The upload must complete within 60 seconds, the file must be at most 25 MiB, and malformed or oversized multipart metadata returns 400 or 413.

Verified-serving constraints

string
A JSON object with the same routing constraints as the JSON provider field, for example {"aci_verified":true}. The header is consumed by the gateway and never forwarded upstream. A provider form field is not accepted.
aci_verified: true requires a verified upstream. aci_session_ids pins a nonempty list of 64-hex session ids and implies verified serving; combining pins with aci_verified: false is invalid. Unavailable pins return 412 session_not_accepted; failed required verification returns 503 upstream_verification_failed, before forwarding the audio. The header must appear once, contain only ASCII bytes, and be at most 8,192 bytes. Malformed or non-object JSON, invalid pins, and unknown aci_ members return 400. JSON endpoints reject this header. Before uploading with constraints, obtain the attestation report over the accepted attested, SPKI-pinned channel, accept its workload provenance, and confirm that service_capabilities.supported_multipart_constraint_endpoints contains /v1/audio/transcriptions. Refuse the upload locally if support is absent. Intersect caller session pins with locally accepted sessions and refuse disjoint sets. Retain the effective policy and check the receipt’s cited session against it: receipt body hashes do not commit to this header. TEE-only host requirements apply independently.

Example

Let the client set the multipart Content-Type boundary. Name Node SDK files explicitly: toFile(buf) uses unknown_file, which has no extension and returns 400.
To require verified serving with cURL, add -H 'X-ACI-Provider: {"aci_verified":true}' after checking the capability above. For Private AI Proxy, auth-headers.txt contains Authorization: Bearer <key>; restrict its permissions to your user and remove it after use. pap curl verifies and pins the gateway channel; it does not audit receipts. Use pap serve or ACI verifier SDK clients for receipt audits (Verify a response); the plain OpenAI SDK examples do not audit receipts.

Response

The cost below is illustrative, not a model price.
json returns text, usage, and optional languages: [{"code":"en"}]. For openai/whisper-large-v3, verbose_json also includes duration and language. Word-level timestamps are not returned. With verbose_json, Whisper responses may include segments; they can be absent. usage contains type: "duration" and integer seconds. Audio pricing is USD per second of input audio, billed per started second: usage.seconds is rounded up (2.965 becomes 3). Verbose duration retains the raw value. The gateway adds usage.cost only when the deployment has audio pricing, in USD: billable seconds times the applicable rate. Provider-reported cost is not echoed. Without a valid duration, the transcript is withheld with 502 and no audio is billed. Read X-Receipt-Id and fetch GET /v1/aci/receipts/{id} using the same bearer token. Audio receipts have chat_id: null, so use the receipt id. See Get receipt and Verify a response.

Errors

Errors use a JSON error object, with error.code and error.param where applicable. Authentication, budget, and other routing constraint errors follow the shared error handling reference.

Models

Find available model ids and capabilities.

Get receipt

Retrieve the signed receipt using X-Receipt-Id.