A transcript tells you what was said; timestamps tell you where in the recording to find it.
How it works
Most ASR systems offer two levels:
- Segment level. One time range per phrase, for example
{"text": " Hi there!", "timestamp": (0.5, 1.5)}. - Word level. One time range per word, for example
hiat 0.5 to 0.9 s andthereat 1.0 to 1.5 s.
Hugging Face’s ASR pipeline exposes both through return_timestamps: True gives segments for Whisper, "word" gives words, and CTC models such as Wav2Vec2 also support "char". Results come back as chunks, each with text and a (start, end) tuple. The docs note that Whisper’s word times are estimated with dynamic time warping over cross-attention weights, an approximation rather than a measurement.
Hosted APIs follow the same pattern. OpenAI’s speech-to-text guide documents timestamp_granularities[] with word or segment, paired with response_format: "verbose_json", and documents it only for the deprecated whisper-1 model, which existing integrations can use until it shuts down on 26 February 2027.
curl https://api.openai.com/v1/audio/transcriptions \
-H "Authorization: Bearer $OPENAI_API_KEY" \
-F model=whisper-1 -F file=@call.mp3 \
-F response_format=verbose_json \
-F "timestamp_granularities[]=word"
Why it matters
- Subtitles. SRT and VTT files need a start and end per caption line.
- Search and review. A quality reviewer can jump to the second a complaint starts instead of listening to a whole call; see transcribing call center recordings.
- Speaker turns. Merging a transcript with diarization output works by matching times.
- Payload size. Word-level output is much longer than the plain text, which matters if you then feed it to an LLM as tokens.
On BroadNet Inference
BroadNet’s docs do not document timestamps. /v1/audio/transcriptions returns the transcript as {"text": "..."}, and no timestamp_granularities option is listed. BroadNet’s /v1/models metadata (checked 2026-11-15) lists verbose_json as a response format, but nothing says it carries segment or word times, so test it before relying on it. For coarse timing, cut the audio yourself, for example at pauses found with voice activity detection or into fixed segments, and record each chunk’s start offset. Each chunk’s text then inherits that offset. This gives sentence-level timing good enough for jumping around a call, not frame-accurate subtitles. Send language with every chunk to BroadNet Transcribe Arabic or BroadNet Transcribe. If word-level timing is a hard requirement, run an aligner or a model that documents it.