How do I get word-level timestamps from speech-to-text?

To get word-level timestamps from speech-to-text, you need an ASR system that supports time alignment for each token/word, not just the full utterance.

Common ways

1) Use an API/model that natively returns word timestamps

Some speech-to-text services can return timestamps per word directly.

Examples:

• OpenAI Whisper API / Whisper-based tools: often provide segment timestamps; some implementations add word-level timing.
• Google Speech-to-Text
• AWS Transcribe
• Azure Speech

Usually you enable an option like:

• word_timestamps=true
• enableWordTimeOffsets
• show_word_time_offsets

2) Use forced alignment after transcription

If your STT model only gives the transcript, you can align the text to audio afterward using a forced aligner.

Popular tools:

• Montreal Forced Aligner
• Gentle
• WhisperX
• aeneas

This is useful when:

• your ASR doesn’t support word timestamps
• you need better accuracy
• you have a transcript already

3) Use Whisper with word-level alignment tools

Whisper itself gives segment timestamps, but tools like WhisperX can add word-level timestamps by:

• running Whisper for transcription
• aligning the transcript to audio with a separate alignment model

This is one of the most common workflows.

If you’re using WhisperX

Typical flow:

1. Transcribe audio
2. Align transcript to audio
3. Read word start/end times

Output often looks like:

[
  {"word": "Hello", "start": 0.52, "end": 0.84},
  {"word": "world", "start": 0.85, "end": 1.12}
]

General advice

To get good word timestamps:

• use clean audio
• make sure the sample rate is standard (often 16 kHz)
• avoid heavy background noise
• if possible, use a model trained for alignment or a dedicated aligner

If you tell me which speech-to-text tool/API you’re using, I can give you the exact steps or code.