#1451·whisperX

Aligner gives inaccurate timestamps for a 25s audio segment (chunk from a larger clip)

Author: bansal-sidCreated Jul 13, 2026Updated Jul 13, 2026

Summary

For a 25-second audio segment, the aligner returned timestamps that are way off from where the speech actually is.

Setup

  • whisperx==3.8.6, CPU
  • Align model: WAV2VEC2_ASR_LARGE_LV60K_960H (torchaudio), language_code="en"
  • This 25-second audio is one chunk from a larger audio file, produced by the transcription step

Input

python
transcript = [{
    "text": "What is it, Timmy?",
    "start": 0.031,
    "end": 25.057,
}]
Expected vs. observed
After running align(), the result comes back at 0.5-1.5
seconds instead — off by roughly 20 seconds.

Repro

import whisperx

device = "cpu"
audio = whisperx.load_audio("clip.wav")  # 25s chunk

transcript = [{
    "text": "What is it, Timmy?",
    "start": 0.031,
    "end": 25.057,
}]

model_a, metadata = whisperx.load_align_model(model_name="WAV2VEC2_ASR_LARGE_LV60K_960H",language_code="en", device=device)
result = whisperx.align(transcript, model_a, metadata, audio, device,
                         return_char_alignments=False)
print(result["segments"][0])

Output

start=0.631, end=1.571

Ask

  • Is it the aligner model failing?
  • I also provided audio by clearing the background noise, but again no success.
  • What can be the possible workarounds to handle such type of cases?

clip_cleaned.wav clip.wav