Results and timestamps
The transcript object, its segments and timestamps, and the warnings that tell you something about the audio.
The result object
GET /v1/transcriptions/{id}/result (or the synchronous response) returns:
| Field | Description |
|---|---|
id | The transcription ID. |
object | transcription.result |
model, version | The model and version that produced it. |
status | completed, or partial when some stretches could not be transcribed. |
mode | The output mode used. |
duration_seconds | Measured length of the audio. |
channels | Channels transcribed: 1, or 2 with channel_mode: "separate". |
languages | Languages in the transcript. |
text | The full transcript as plain text. |
segments | Ordered segments — see below. |
warnings | Codes describing the audio or result — see below. |
gaps | Only when status is partial: the {start, end} stretches that are missing. |
Segments
| Field | Description |
|---|---|
id | Position in the transcript, from 0. |
start, end | Seconds from the start of the recording. Omitted with timestamps: "none". |
text | The words of this segment. |
speaker | With diarization: true: speaker_1, speaker_2, … |
overlap | With diarization: true where speakers overlap. |
channel | With channel_mode: "separate": 0 or 1. |
language | The segment's language. |
Timestamps
timestamps: "segment" (the default) gives every segment its start and end time. In the 22 Indian languages segments are sentence-sized; English segments can be longer. Word-level timestamps (timestamps: "word") are coming soon.
Plain text
Ask for GET /v1/transcriptions/{id}/result?format=txt to get the transcript as plain text, or read the text field of the JSON result. SRT and WebVTT output are coming soon (format=srt / vtt return capability_unavailable); until then the subtitles recipe builds both from segments in a few lines.
Progress
While a transcription runs, GET /v1/transcriptions/{id} includes progress — the percentage of the audio already transcribed. It is 0 before work starts, a measured value while transcribing (null until the first measurement arrives — never an estimate), and 100 once the result exists.
Warnings
Warnings never change the shape of the result; they tell you something worth knowing.
| Warning | Meaning |
|---|---|
no_speech_detected | No speech was found; text is empty. |
partial_result | Some stretches could not be transcribed; they are listed in gaps. |
language_low_confidence | The audio may not be in the requested language. |
audio_resampled | The audio was converted to the sample rate used for transcription. |
low_sample_rate | The recording’s sample rate is low; accuracy may suffer. |
channels_mixed | A multi-channel recording was mixed to one channel. |
short_audio_padded | Very short audio was padded before transcription. |
word_timestamps_partial | Some words have no timing. |
possible_repetition | A segment still repeats itself; its text is returned as produced, never cut short. |
repetition_recovered | A segment that repeated itself was transcribed again more plainly; mode formatting (English in Latin script, fillers) may not apply to it. |
speaker_labels_coarse | Some segments carry one speaker label for their whole span. |
Partial results
A partial result is returned exactly like a completed one, with the partial_result warning and the missing stretches in gaps. Partial results are not charged. To try again, start a new transcription with a new idempotency key.