# What are word-level timestamps, and which APIs return them?

A word-level timestamp is the start and end time of one spoken word. See which transcription and text to speech APIs return them, what they are for, and the cost.

Published 2026-11-17. Paxa Labs.

A word-level timestamp is the start and end time of one spoken word in an audio file. For a recording, Paxa STT Lite returns one for every word in the transcript. For generated speech, Paxa TTS Flash v2 returns them with the audio. Deepgram, AssemblyAI, ElevenLabs, Amazon Transcribe, Google, and Soniox document word timing for transcripts, and Cartesia, Microsoft Azure, Amazon Polly, and WellSaid for speech.

## Which transcription APIs give a timestamp for every word spoken?

Each row links the vendor's own documentation and restates only what that page says. Vendors appear in no order of preference.

| API | How the documentation names it | What you get |
| --- | --- | --- |
| [Paxa STT Lite](https://paxalabs.com/docs/speech-to-text#timestamps) | `timestamps` set to `word` | A `words` array, one whole word per entry with its start and end in seconds. Thai is segmented into words. |
| [Deepgram](https://developers.deepgram.com/docs/pre-recorded-audio) | A `words` array on a pre-recorded response | Each word with `start` and `end` in seconds from the beginning of the audio. |
| [AssemblyAI](https://www.assemblyai.com/docs/api-reference/transcripts/get) | A `words` array on the transcript | One object per word in the transcript, each with `start` and `end`. |
| [ElevenLabs Scribe](https://elevenlabs.io/docs/api-reference/speech-to-text/convert) | `timestamps_granularity`, set to `word` or `character` | Each word with `start` and `end` in seconds, or the characters inside each word. |
| [Amazon Transcribe](https://docs.aws.amazon.com/transcribe/latest/dg/how-input.html) | An `items` list in the transcript JSON | A start time and an end time for every word. |
| [Google Chirp 3](https://docs.cloud.google.com/speech-to-text/v2/docs/chirp_3-model) | Word-level timestamps on BatchRecognize | Audio of up to 20 minutes when they are enabled, and up to an hour without them. |
| [OpenAI](https://developers.openai.com/api/docs/guides/speech-to-text) | `timestamp_granularities` | Word or segment timestamps, supported on whisper-1 only. |
| [Soniox](https://soniox.com/docs/stt/concepts/timestamps) | Token timestamps | `start_ms` and `end_ms` on every token. A token is a whole word or a piece of one. |
| [Speechmatics](https://docs.speechmatics.com/speech-to-text/models) | Word timings | Listed for its Batch models. |

## Which text to speech APIs return word timestamps with the audio?

Generated speech needs timing for captions and read-along highlighting, and the vendors deliver it in different shapes. Some return it in the same response as the audio. Others send it as events or as a separate file.

| API | How the documentation names it | What you get |
| --- | --- | --- |
| [Paxa TTS Flash v2](https://paxalabs.com/docs/timestamps) | `timestamps` set to `word`, `sentence`, or `utterance` | Spans with their text, start, and end in seconds, in the same response as the audio. Thai text needs no spaces. |
| [ElevenLabs](https://elevenlabs.io/docs/api-reference/text-to-speech/convert-with-timestamps) | A with-timestamps endpoint | The audio as base64 with a start and end time for each character. |
| [Cartesia](https://docs.cartesia.ai/api-reference/tts/tts) | `add_timestamps` | Word-level timing returned with the audio. |
| [Microsoft Azure](https://learn.microsoft.com/en-us/azure/ai-services/speech-service/how-to-speech-synthesis) | The `WordBoundary` event in the Speech SDK | The time offset of each spoken word, punctuation mark, and sentence. |
| [Amazon Polly](https://docs.aws.amazon.com/polly/latest/dg/speechmarks.html) | Speech marks | Metadata marking where sentences and words start and end. A speech marks request returns the marks, and the audio comes from its own request. |
| [WellSaid](https://docs.wellsaidlabs.com/reference/word-timing) | A word timing endpoint | A zip file with the audio and word timing as JSON, SRT, and VTT. |

## What does a word-level timestamp look like?

Each entry holds three things: the word as text, the second it starts, and the second it ends. On Paxa STT Lite the entries arrive as a `words` array beside the transcript, timed on the recording's own clock. A punctuation mark travels with the word next to it. Joining the entries back together gives the transcript with nothing lost.

On Paxa TTS Flash v2 the same shape is called a span. A request can ask for one span per word, per sentence, or per run of speech between pauses. [The timestamps guide](https://paxalabs.com/docs/timestamps) shows a full response, and [the transcription guide](https://paxalabs.com/docs/speech-to-text#timestamps) shows the array for a recording.

## What do people build with word timestamps?

- Captions and subtitles. Each caption line starts and ends on a word. The text on screen then matches the voice. [The subtitle generator](https://paxalabs.com/subtitle-generator) returns an SRT or WebVTT file built this way.
- Read-along highlighting. A lesson, an article, or a karaoke line lights each word as it is spoken. [The subtitles page](https://paxalabs.com/subtitles) plays an example.
- Search inside audio. A search for a phrase in a meeting or a podcast jumps to the second it was said.
- Editing by text. Deleting a sentence from the transcript gives the cut points for the same sentence in the audio.
- Lip sync and animation. A character's mouth, a slide change, or a subtitle effect follows the timing of each word.

## Why do Thai and Chinese need the API to decide what a word is?

English puts a space between words. Everyone agrees where one word ends. Thai and Chinese are written with no spaces between words. An API that times such text has to decide where each word ends before it can give the word a time. Some return a time for each character or each token, and the developer then has to join the pieces into words.

Paxa STT Lite segments Thai for you, and each entry covers a whole word. English words and numerals inside a Thai sentence arrive whole too. Subtitle lines break on those word boundaries, which keeps a Thai word from being split across two lines. [Thai subtitles from word timestamps](https://paxalabs.com/guides/thai-subtitles-from-word-timestamps) walks through a caption file. For audio that mixes languages, see [the speech to text APIs for multilingual audio](https://paxalabs.com/guides/best-speech-to-text-api-for-multilingual-audio).

## How much do word timestamps cost?

On transcription they cost nothing extra. Paxa STT Lite bills 8.33 credits per minute of audio, $0.50 per hour, with or without word timestamps, speaker numbers, or a subtitle file. On speech, a request that asks for timestamps bills at 1.25 times the speech rate. Paxa TTS Flash v2 costs 10 credits per 1,000 characters, $10 per million, before that multiplier.

A new account gets 100 free credits with no card, which cover about 12 minutes of transcription. Vendor rates are on [the speech to text API comparison](https://paxalabs.com/compare/speech-to-text-apis) and [the text to speech API comparison](https://paxalabs.com/compare/text-to-speech-apis).

## How do you try word timestamps without writing code?

- Upload a recording to [the Transcribe tool](https://paxalabs.com/app/transcribe) and turn on word timestamps. The result lists each word with its time, and a subtitle file downloads from the same screen.
- Paste a script into [the Speech tool](https://paxalabs.com/app/speech) and turn on timestamps. The words light up as the voice reads them.

## What should you read next?

- [Get word-level timestamps](https://paxalabs.com/docs/timestamps) for the speech request and response.
- [Transcribe recordings](https://paxalabs.com/docs/speech-to-text) for the transcription options.
- [The subtitle generator](https://paxalabs.com/subtitle-generator) for SRT and WebVTT files from a recording.
- [The text to speech APIs for voice agents](https://paxalabs.com/guides/best-text-to-speech-api-for-voice-agents) for timing on a live connection.

## Common questions

### Which API gives a timestamp for every word spoken?

For recordings, Paxa STT Lite returns a start and an end for every word when `timestamps` is set to `word`. Deepgram, AssemblyAI, ElevenLabs Scribe, and Amazon Transcribe document a per-word start and end as well. For generated speech, Paxa TTS Flash v2 and Cartesia return word timing with the audio.

### How much does it cost?

Word timestamps on a transcript are included in the 8.33 credits per minute that Paxa STT Lite bills. On Paxa TTS Flash v2, a timed request bills at 1.25 times the speech rate.

### Does OpenAI return word-level timestamps?

[OpenAI's transcription guide](https://developers.openai.com/api/docs/guides/speech-to-text) supports `timestamp_granularities` on whisper-1 only, and [its realtime transcription guide](https://developers.openai.com/api/docs/guides/realtime-transcription) states that gpt-live-transcribe returns no word-level timestamps. Paxa's OpenAI-compatible transcription endpoint returns words when `timestamp_granularities` names `word`, and [live speech connections](https://paxalabs.com/docs/live-streaming) carry timing too.

### Do word timestamps work on Thai?

Yes. Paxa STT Lite segments Thai into words before timing them, and Paxa TTS Flash v2 does the same for the text it reads. No spaces need to be added to the text.

### Do I need word timestamps to get a subtitle file?

No. Ask Paxa STT Lite for an SRT or WebVTT file and it reads the word timing itself. Asking for both returns the file and the word list in one response, at no extra cost.
