AI Video Editing Flows

We tested faster-whisper captions on three dialogue clips

faster-whisper turns speech into timed text you can use for captions. We ran two of its English models on three 60-second clips from the open film Tears of Steel on a laptop CPU and compared the words with the film's official subtitles.

Tested by us on Nov 5, 2026, on the clips and machine listed below. Every number on this page was measured in that run.

The result

The small English model (small.en) got closer to the official subtitles than the base model on all three clips: a word error rate of 10.9% to 17.0%, against 19.0% to 26.6% for base.en. On a laptop CPU, small.en took a median of 30 to 35 seconds per 60-second clip, and base.en 8 to 41 seconds, with the laptop busy with other work.

Read those error rates with care: a good part of them is not a transcription mistake. The subtitles write "ten minutes" where the model writes "10 minutes", spell a name "Thom" where it writes "Tom", and shorten some lines. The clearest real mistake we found: "We're on in one" came out as "We're only one".

Neither model filled the quiet stretches with invented sentences: each placed just one word at a moment the subtitles show no line, both on the third clip ("Whoa!" from base.en, "I'm" from small.en).

What failed or surprised us

  • faster-whisper 1.2.1 could not open any of our files with the current PyAV (19.0.1 and 19.0.2): every run stopped with a TypeError about metadata_errors. We decoded the audio with FFmpeg and passed it in as samples, which works and is included in every timing.
  • The first run of small.en took 175 seconds, because it downloaded the model. Later runs took about 30 seconds.
  • Timings on a shared laptop swing a lot: the same base.en job took 8 seconds on one clip and 41 on another of the same length.
  • Numbers, names and contractions count as errors against subtitles even when the words are right, so the raw error rate overstates real mistakes.

What we tested and how

  • Tool: faster-whisper version 1.2.1, MIT licence.
  • Settings compared: Model base.en, int8 on CPU; Model small.en, int8 on CPU.
  • Input clips: Tears of Steel, (CC) Blender Foundation, mango.blender.org, CC-BY 3.0. 60 s excerpts re-encoded to H.264 720p.
  • tos_a: 60 s starting at 20 s of the film, SHA-256 945262dcbe106f1e...
  • tos_b: 60 s starting at 180 s of the film, SHA-256 4097773b615cddfa...
  • tos_c: 60 s starting at 360 s of the film, SHA-256 3d3a8cd015310349...
  • Machine: AMD Ryzen 7 7735U (8 cores, 16 threads), integrated Radeon 680M, not used (CPU-only runs), 31.2 GB RAM, Windows 11 10.0.26300. The laptop was shared with other work, so timings are what a busy machine gives, not a clean benchmark.
  • Runs: one cold run, then three timed runs per clip and setting. Run on 2026-11-05.

faster-whisper transcribes speech into timed words, the raw material for captions. We ran two of its English models on the CPU (int8, 8 threads, beam size 5, word timestamps on, no voice-activity filter) on the same three clips, then compared each transcript with the film's official English subtitles for the same 60 seconds.

Measurements by clip

Model base.en, int8 on CPU

ClipWord error rate vs subtitlesWords (subtitles / ours)Words outside subtitle linesMedian run (s)Real-time factorPeak memory (MB)
tos_a19%79 / 7707.80.131343
tos_b26.6%64 / 57040.50.675426
tos_c26.4%53 / 51119.80.331316

Model small.en, int8 on CPU

ClipWord error rate vs subtitlesWords (subtitles / ours)Words outside subtitle linesMedian run (s)Real-time factorPeak memory (MB)
tos_a15.2%79 / 76029.70.496675
tos_b10.9%64 / 63034.70.579676
tos_c17%53 / 51135.40.59658

Reproduce it

Model base.en, int8 on CPU
import json, sys, os, subprocess
import numpy as np
from faster_whisper import WhisperModel
FFMPEG = "ffmpeg"
raw = subprocess.run([FFMPEG, "-v", "error", "-i", sys.argv[2], "-ac", "1", "-ar", "16000", "-f", "f32le", "-"], capture_output=True, check=True).stdout
audio = np.frombuffer(raw, dtype=np.float32)
m = WhisperModel(sys.argv[1], device="cpu", compute_type="int8", cpu_threads=8)
segs, info = m.transcribe(audio, language="en", word_timestamps=True, vad_filter=False, beam_size=5)
out = {"language": info.language, "duration": info.duration, "segments": []}
for s in segs:
    out["segments"].append({"start": s.start, "end": s.end, "text": s.text, "words": [{"w": w.word, "s": w.start, "e": w.end, "p": w.probability} for w in (s.words or [])]})
json.dump(out, open(sys.argv[3], "w", encoding="utf-8"), ensure_ascii=False)

# run: python transcribe.py small.en clip.mp4 out.json
Model small.en, int8 on CPU
import json, sys, os, subprocess
import numpy as np
from faster_whisper import WhisperModel
FFMPEG = "ffmpeg"
raw = subprocess.run([FFMPEG, "-v", "error", "-i", sys.argv[2], "-ac", "1", "-ar", "16000", "-f", "f32le", "-"], capture_output=True, check=True).stdout
audio = np.frombuffer(raw, dtype=np.float32)
m = WhisperModel(sys.argv[1], device="cpu", compute_type="int8", cpu_threads=8)
segs, info = m.transcribe(audio, language="en", word_timestamps=True, vad_filter=False, beam_size=5)
out = {"language": info.language, "duration": info.duration, "segments": []}
for s in segs:
    out["segments"].append({"start": s.start, "end": s.end, "text": s.text, "words": [{"w": w.word, "s": w.start, "e": w.end, "p": w.probability} for w in (s.words or [])]})
json.dump(out, open(sys.argv[3], "w", encoding="utf-8"), ensure_ascii=False)

# run: python transcribe.py small.en clip.mp4 out.json

Install faster-whisper with pip. Decode the clip's audio to 16 kHz mono samples with FFmpeg, then pass the samples to the model. This is the exact script we ran for each clip, with the model name, the clip and the output file as arguments.

Method and limits

  • Word error rate is measured against the official English subtitles after lower-casing and removing punctuation. Subtitles can shorten speech and write numbers as words, so the figure likely overstates real mistakes; it is a comparison between the two models, not an exact score.
  • Three 60-second clips of acted film dialogue with music underneath. A clean podcast or talking-head recording would likely score better; a noisy street interview worse.
  • CPU only, on a laptop that was running other work; the CPU load just before each run is stored with the results. A GPU would be much faster.
  • We did not test larger models, other languages, or the voice-activity filter yet; those are next.
  • Footage: Tears of Steel, (CC) Blender Foundation, mango.blender.org, used under CC-BY 3.0. We publish the measurements, not the transcripts as captions.

Page checked on Nov 4, 2026.