Skip to Content
API AccessTimed speech transcription

Audio transcription

POST /v1/audio/transcriptions accepts the standard OpenAI-compatible multipart/form-data upload. Configure an OpenAI Compatible or OpenAI connection and select its concrete transcription model.

curl "$STEINKAUZ_ORIGIN/v1/audio/transcriptions" \
  -H "Authorization: Bearer $STEINKAUZ_API_KEY" \
  -H "X-Steinkauz-Data-Sensitivity: D2_CONFIDENTIAL" \
  -F 'model=connection/CONNECTION_UUID/MODEL_ID' \
  -F 'file=@prepared-audio.wav' \
  -F 'response_format=verbose_json' \
  -F 'timestamp_granularities[]=word' \
  -F 'timestamp_granularities[]=segment'

Supported fields are file, model, optional language, prompt, temperature (0–1), response_format and repeated timestamp_granularities[]. Formats are json (default), verbose_json, text, srt and vtt, subject to upstream support. Timestamp granularities word and segment require verbose_json. Streaming and diarized output are not implemented in this version and fail explicitly. JSON audio/base64 and custom source-coordinate fields are not accepted.

Ordinary JSON output contains text. Verbose output additionally reports language, duration, words with word/start/end, and/or segments with text/start/end. Steinkauz validates requested timing arrays, finite chronological ranges and bounds against the reported duration. Times remain relative to the uploaded clip. The consuming application owns source-duration completeness checks, chunking and conversion to track offsets. No custom complete or source fields are added.

The selected service transcribes the audio using the requested model and multipart fields. Requested word and segment timings are preserved. If the service omits requested timings, the request fails; another connection is not tried.

Uploads are limited to 25 MiB plus bounded multipart overhead. Upstream services may impose lower limits. Deadlines and cancellation propagate. Authentication, classification, D/E authorization, budget checks and required evidence precede provider egress. Audio bodies are omitted from retained content and replaced by observed size/digest evidence. Text and transcript output follow content retention. Reported input/output token usage is recorded when available; duration-only usage does not invent token counts or a cost.

See the OpenAI transcription reference  for the standard formats and model-specific capabilities.

EmbeddingsDecisions