How do I generate timestamps from audio transcription?

To generate timestamps from audio transcription, you usually need to align the transcript text with the audio. There are a few common ways to do it, depending on what tools you have.

1) Use a speech-to-text model that already outputs timestamps

Many ASR tools can return:

• segment timestamps: start/end time for each chunk of speech
• word timestamps: time for each individual word

Examples:

• OpenAI Whisper / faster-whisper
• Google Speech-to-Text
• AWS Transcribe
• Azure Speech
• AssemblyAI

If you’re using one of these, timestamp generation is often built in.

Example with Whisper-style output

You may get output like:

[
  {
    "start": 0.0,
    "end": 4.2,
    "text": "Hello everyone, welcome to the show."
  },
  {
    "start": 4.2,
    "end": 8.1,
    "text": "Today we'll talk about audio transcription."
  }
]

2) If you only have plain text transcript

You can force alignment by comparing the transcript to the audio.

Common alignment approaches:

• Word-level forced alignment: aligns known transcript words to audio
• Phoneme alignment: more precise, often used in research or advanced tools

Tools for forced alignment:

• WhisperX
• Montreal Forced Aligner
• Gentle
• aeneas

These tools take:

• audio file
• transcript text

and return:

• word timestamps
• sentence timestamps

3) Generate timestamps by splitting audio into chunks

If you don’t need exact timestamps, you can:

1. split audio into fixed-length chunks, e.g. 10 or 30 seconds
2. transcribe each chunk separately
3. assign timestamps based on chunk boundaries

This is easier, but less accurate.

4) Best practice: use word timestamps, then build sentence timestamps

If you have word-level times:

• group words into sentences
• sentence start = first word start
• sentence end = last word end

Example:

0.0 Hello everyone,
1.2 welcome to the show.

Sentence timestamp:

• start: 0.0
• end: 1.2

5) Simple Python example with WhisperX

import whisperx

device = "cuda"  # or "cpu"
audio_file = "audio.mp3"
model = whisperx.load_model("medium", device)

result = model.transcribe(audio_file)

align_model, metadata = whisperx.load_align_model(
    language_code=result["language"],
    device=device
)

aligned = whisperx.align(
    result["segments"],
    align_model,
    metadata,
    audio_file,
    device
)

for seg in aligned["segments"]:
    print(seg["start"], seg["end"], seg["text"])

This can give you timestamps at the word level too.

6) Output formats you can generate

You can convert timestamps into:

• SRT
• VTT
• JSON
• CSV

Example SRT

1
00:00:00,000 --> 00:00:04,200
Hello everyone, welcome to the show.

2
00:00:04,200 --> 00:00:08,100
Today we'll talk about audio transcription.

7) What to choose?

• Want easy and good accuracy: use Whisper with timestamps
• Want word-level precision: use WhisperX or forced alignment
• Want simple rough timestamps: split audio into chunks

If you want, I can also show you:

1. how to do this in Python,
2. how to convert transcripts into SRT/VTT, or
3. how to get word-level timestamps with Whisper.