The short answer
Word timestamps are the start and end times, in seconds from the beginning of the audio, that a speech recognition system attaches to each word of its transcript. They are what subtitle files, click-to-seek transcripts, audio redaction and speaker-labelled transcripts are built from.
Phonon-2.5, Phonon-2.5 English and Phonon-2 return them with fermion transcribe --json, and the server returns them in the OpenAI verbose_json shape.
Three ways to get word timings
Where the times come from depends on the model family.
| Method | Where the times come from | Extra model needed | Public example |
|---|---|---|---|
| From the decoder's own alignment | CTC and transducer models emit tokens at specific audio frames, so each word's frames give its times | No | Parakeet TDT 0.6B v3 lists word-level timestamps on its card |
| Cross-attention plus dynamic time warping | The decoder's attention weights over audio frames are smoothed and aligned to tokens | No, but timing is indirect | Whisper's word_timestamps option (whisper/timing.py) |
| Forced alignment after transcription | A separate acoustic model aligns the finished text to the audio | Yes | WhisperX: voice activity detection plus forced phoneme alignment (Bain et al., 2023) |
Getting word timestamps from Phonon
On the command line, --json returns a JSON object with the text, a start and end time per segment, and a words list with the start and end of every word in seconds from the start of the file.
On the server started with fermion serve, POST /v1/audio/transcriptions with response_format=verbose_json and timestamp_granularities=word returns a words list (word, start, end) and a segments list in the shape OpenAI clients expect. Uploads go up to 1 GB; the model handles utterances up to 30 seconds and decodes longer audio in chunks.
fermion transcribe phonon-2.5 clip.wav --json
curl -s http://127.0.0.1:8000/v1/audio/transcriptions -F file=@clip.wav -F response_format=verbose_json -F timestamp_granularities=wordWhat affects timing accuracy
- Frame resolution: a model's encoder looks at audio in fixed steps, so a word boundary cannot be placed more finely than one step.
- Fast or overlapping speech: when words run together, boundaries are ambiguous even for human annotators.
- Pauses and punctuation: some systems attach a pause to the previous word's end, others to the next word's start; check which before cutting audio.
- Long recordings: audio split into chunks needs each chunk's offset added back, which a well-built tool does for you.
To check timings on your own audio, compare them with a hand-marked or forced-aligned reference and measure the average gap at word starts and ends.
Questions
- What are word timestamps?
- The start and end time of each word in a transcript, in seconds from the start of the audio. They are used for subtitles, click-to-seek transcripts and redaction.
- How do I get word timestamps from Phonon?
- Run fermion transcribe phonon-2.5 file.wav --json, which includes a words list with start and end times. On the server, use response_format=verbose_json with timestamp_granularities=word.
- Which Phonon models return word timestamps?
- Phonon-2.5, Phonon-2.5 English and Phonon-2. Phonon-1 returns null for the words list.
- How does Whisper produce word timestamps?
- Its word_timestamps option aligns tokens to audio frames from selected cross-attention heads, smoothed with a median filter and aligned with dynamic time warping.
- What is forced alignment?
- Aligning a known transcript to audio with an acoustic model after transcription. WhisperX, for example, combines voice activity detection with forced phoneme alignment to time each word.
- Are word timestamps the same as speaker labels?
- No. Timestamps say when each word was spoken; speaker labels (diarization) say who spoke it. Diarization is a separate step that can use word timings to assign words to speakers.
Related
Sources and method
- Phonon command line and server behaviour: Fermion Research docs (/docs/cli-reference/, /docs/server-reference/), read 2026-11-13.
- Whisper word timing: github.com/openai/whisper, whisper/timing.py, read 2026-11-13.
- Bain, Huh, Han, Zisserman, WhisperX: Time-Accurate Speech Transcription of Long-Form Audio, arXiv 2303.00747 (2023-03-01), read 2026-11-13.
- Parakeet TDT 0.6B v3 model card: huggingface.co/nvidia/parakeet-tdt-0.6b-v3, read 2026-11-13.
Canonical: https://www.fermionresearch.com/glossary/word-timestamps/