Audio Transcription API: Structured Transcripts, Not Just Speech-to-Text

Most teams treat audio transcription as a convenience feature: press a button, get a wall of text, skim it, move on. That framing stops being useful the moment the transcript has to do something. Feeding a search index, lining up against an audio player, triggering a review, or landing in a compliance record all need more than a block of words, and plain text can't carry any of it on its own.

Taareef treats transcription as a structured data layer, not a speech-to-text toy. Every recording (a call, a meeting, a voice note) comes back as JSON: the full transcript, the detected language, the clip's duration, and a breakdown into timed segments and individual words, each carrying its own confidence score. This guide walks through exactly what the API returns and why that structure is what makes a transcript usable in a real pipeline.

What does audio transcription return?

Audio transcription returns a structured JSON object, not a plain string. A single POST of an audio file comes back with the complete transcription text, the auto-detected language, the duration in seconds, an overall confidence score, and a segments array that breaks the audio into timed chunks. Each segment carries its own start and end timestamps, its text, a segment-level confidence, and a words array in which every word has its own timing and confidence.

{  "transcription": "The new prototype is ready for internal review by next Tuesday.",  "language": "en",  "duration": 6.718,  "confidence": 0.9803,  "segments": [    {      "start": 0.0,      "end": 3.54,      "text": "The new prototype is ready for internal review",      "confidence": 0.9876,      "words": [ … ]    }    { … }  ]}

Why plain text isn't enough

A raw transcript answers what was said and nothing else. It doesn't tell you when each phrase happened, how confident the model was about any given word, or which language it was in. Anything you build on top (jumping the audio player to a quoted line, flagging a mumbled passage for review, indexing the call for search) needs that metadata, so with plain text you end up re-deriving it in a fragile second pass that is slow and error-prone. Returning it in the first response removes that step entirely.

Word-level and segment-level timestamps

Every segment and every word comes back with start and end times in seconds. That timing is what turns a transcript into something interactive: click a line and jump the player to that exact moment, generate captions that line up with the audio, clip the segment where a phrase was spoken, or locate an exact line for dispute resolution and audit trails without re-listening to the whole recording.

Confidence scores you can gate on

The API reports a confidence score between 0 and 1 at three levels: one for the whole recording, one per segment, and one per word. A low score is a signal, not a failure. It usually marks a mumbled word, crosstalk, or background noise. Teams use it to route uncertain passages to a human reviewer, suppress low-confidence words in a caption, or reject a recording that is too noisy to trust, instead of treating every transcript as equally reliable.

Automatic language detection

Send language=auto and the API identifies the spoken language, returning its code (for example en) alongside the transcript, across 40+ languages. One endpoint then handles a multilingual call center or a mixed-language user base without the caller having to declare a language up front, and the detected code travels with the result so downstream steps can route on it.

Built for developers, and for the rest of the pipeline

Audio transcription is a developer-first API: send an audio file in any common format, get back clean, predictable JSON, no manual cleanup in between. Because it uses the same structured-JSON shape as the rest of Taareef, a voice step drops into an identity or compliance flow without a translation layer between vendors. A KYC call, a recorded consent, or a support interaction becomes just another structured input next to the Emirates ID read and the selfie-to-ID face match.

Where structured transcription fits

• Call centers and support: searchable, timestamped records of every conversation, with low-confidence passages flagged for QA.
• Meetings and voice notes: turn recordings into text your app can index, quote, and link back to the exact moment.
• KYC and recorded consent: capture what was said and when, as a structured record that sits alongside document and face checks.
• Captioning and accessibility: word-level timing drives subtitles that track the audio.
• Compliance and audit: timestamped, confidence-scored transcripts you can point to later without re-listening.

Why it matters

Speech-to-text tells you what was said. A structured transcript tells you what was said, when, in which language, and how reliably, in a shape your code can act on immediately. That is the difference between a transcript you read and a transcript you can build on, and it runs through the same JSON and the same single API as the rest of your identity and document checks.