Transcribe - Framelane

Overview

Transcribes audio or video using AssemblyAI Universal-2/3 models. Returns a structured transcript with optional word timings, speaker labels, sentiment, entities, and content safety flags.

Request

POST /v1/tasks/transcribe
Authorization: Bearer {api_key}

{
  "source_url": "https://cdn.example.com/interview.mp4",
  "language": "en",
  "speech_model": "universal-2",
  "word_timings": true,
  "speaker_labels": true,
  "speakers_expected": 2
}

Key parameters

Parameter	Default	Description
`source_url`	required	URL of the audio or video file
`language`	auto	BCP-47 language code, e.g. `en`, `de`, `es`
`speech_model`	`universal-2`	`universal-2`, `universal-3-pro`, or `nano`
`word_timings`	`true`	Include word-level timestamps
`speaker_labels`	`false`	Enable speaker diarisation
`speakers_expected`	null	Hint for number of speakers (1–10)
`sentiment_analysis`	`false`	Per-sentence sentiment
`entity_detection`	`false`	Named entity recognition
`auto_highlights`	`false`	Extract key phrases
`content_safety`	`false`	Flag sensitive content
`iab_categories`	`false`	IAB taxonomy classification
`word_boost`	`[]`	Domain terms to boost (max 200)

Output

Transcribe returns data, not a file. When status is completed, the transcript is on the result object of the job itself — GET /v1/tasks/{id} (or the 202 create response, if the provider already finished). There is nothing to download: output is null, and GET /v1/tasks/{id}/download answers 404 for a transcribe task.

curl -H "Authorization: Bearer $FRAMELANE_API_KEY" \
  "https://api.framelane.io/v1/tasks/task_01J8QR2K5VKDGN2T4FBM3CZYX7?wait=60" \
  | jq '.result.text'

result holds the normalised AssemblyAI response. words and utterances are always present (empty arrays when the provider returned none); the analysis blocks appear only for the features you enabled.

Field	Description
`text`	Full transcript text
`language_code`	Detected or requested language
`confidence`	Overall confidence (0–1)
`audio_duration`	Source duration in seconds
`words[]`	`text`, `start`, `end` (ms), `confidence`, `speaker`
`utterances[]`	Speaker-diarised turns — populated when `speaker_labels` is on
`sentiment_analysis[]`	Present when `sentiment_analysis` is on
`entities[]`	Present when `entity_detection` is on
`auto_highlights[]`	Present when `auto_highlights` is on
`content_safety`	Present when `content_safety` is on
`iab_categories`	Present when `iab_categories` is on

Example completed task

{
  "id": "task_01J8QR2K5VKDGN2T4FBM3CZYX7",
  "kind": "task",
  "task_type": "transcribe",
  "status": "completed",
  "output": null,
  "result": {
    "text": "Welcome back to the show...",
    "language_code": "en",
    "confidence": 0.97,
    "audio_duration": 812,
    "words": [
      { "text": "Welcome", "start": 0, "end": 650, "confidence": 0.99, "speaker": "A" }
    ],
    "utterances": [
      { "speaker": "A", "text": "Welcome back to the show.", "start": 0, "end": 2500 }
    ]
  }
}

Values are passed through from AssemblyAI unchanged, including their units: word and utterance start / end are milliseconds, while audio_duration is in seconds. word_animation in a render request takes seconds — divide the word timings by 1000 when feeding them into a composition.

The task.completed webhook carries status and metadata only, not the transcript — fetch GET /v1/tasks/{id} when it fires.