FFmpeg 8.0 added a whisper filter that turns speech into text. It runs whisper.cpp, so with one model file you can make SRT subtitles from a video in a single command.

Measured on Japanese and English readings, the setting that mattered most was not the model but queue, the length of audio processed at a time:

  • With the default queue (3 seconds), large-v3-turbo on the GPU got 27% of the Japanese and 12% of the English wrong. With queue=20 that fell to 12% and 3%, and it ran faster too.
  • On the GPU, adding VAD (detection of the parts without speech) made it more than 10× slower.

Basic Command

This makes SRT subtitles from a video.

queue=20 is the setting that matters. Use the ggml-large-v3-turbo-q5_0.bin model (download below).

ffmpeg -i input.mp4 -vn -af "whisper=model=ggml-large-v3-turbo-q5_0.bin:language=en:queue=20:destination=output.srt:format=srt" -f null -
Option Meaning
model= The model file
language=en The spoken language. The default, auto, detects it
queue=20 Seconds of audio processed at a time. Default 3
destination= The file to write
format=srt Output format: text (text only), srt or json
use_gpu Whether to use the GPU. On by default (use_gpu=0 for CPU only)

-f null - writes no audio. The SRT numbers its cues from 0, but they are renumbered when you add it to a video, so use it as is.

Writing Paths on Windows

Options are separated by :, so an absolute path such as C:\... gets cut in two. Use one of these:

# 1. Put the model and the video in one folder, cd there, and use file names only (output.srt is written there too)
cd C:\work
ffmpeg -i input.mp4 -vn -af "whisper=model=ggml-large-v3-turbo-q5_0.bin:language=en:queue=20:destination=output.srt:format=srt" -f null -

# 2. Wrap the value in ' and put \ before the : (use / as the separator)
ffmpeg -i input.mp4 -vn -af "whisper=model='C\:/models/ggml-large-v3-turbo-q5_0.bin':language=en:queue=20:destination=output.srt:format=srt" -f null -

Writing just C\: without the ' quotes failed with Invalid argument.

Check That Your FFmpeg Has It

The whisper filter exists only in FFmpeg 8.0 and later builds that include whisper.cpp. gyan.dev’s full build has it; the essentials build does not (gyan.dev’s list).

ffmpeg -hide_banner -h filter=whisper

Unknown filter 'whisper' means your build does not include it.

Get a Model

Download one whisper.cpp model (ggml-*.bin) from Hugging Face.

Model File Size
base ggml-base.bin 148 MB
small ggml-small.bin 488 MB
large-v3-turbo (quantized) ggml-large-v3-turbo-q5_0.bin 574 MB
large-v3-turbo ggml-large-v3-turbo.bin 1.62 GB

If unsure, pick ggml-large-v3-turbo-q5_0.bin: in the measurements below it was about as accurate as the unquantized model at a third of the size.

Add the Subtitles to the Video

This adds the SRT as MP4 subtitles (mov_text) and copies the video and audio unchanged:

ffmpeg -i input.mp4 -i output.srt -map 0 -map 1 -c copy -c:s mov_text -metadata:s:s:0 language=eng output_sub.mp4

For burning subtitles into the picture and the differences between subtitle formats, see the subtitles article.

Measured: How Model and Settings Change Accuracy and Speed

Error rates were measured on readings of Gongitsune in Japanese (15:46) and The Gift of the Magi in English (13:22).

Queue Length (large-v3-turbo, GPU)

Setting Japanese: time / errors English: time / errors
Default (3 s) 73 s / 26.8% 46 s / 11.9%
queue=20 56 s / 12.0% 33 s / 3.0%
queue=20 + VAD 753 s / 12.5% 872 s / 3.0%

VAD is enabled by passing a Silero VAD model (ggml-silero-v5.1.2.bin, Hugging Face) as vad_model=.

Models Compared (GPU, queue=20)

Model Japanese: time / errors English: time / errors
base 71 s / 22.2% 39 s / 6.3%
small 60 s / 17.3% 40 s / 9.6%
large-v3-turbo (quantized q5_0) 51 s / 12.5% 36 s / 3.2%
large-v3-turbo 56 s / 12.0% 33 s / 3.0%

small dropped several English sentences near the start (with VAD it did not, and its error rate fell to 3.9%).

CPU Only (queue=20, use_gpu=0)

Model Japanese: time / errors English: time / errors
small 262 s / 17.4% 190 s / 9.6%
large-v3-turbo (quantized q5_0) 346 s / 11.6% 252 s / 2.9%

Things to Check in a Transcript

  • Names: even large-v3-turbo wrote “Madame Sofronie” as “Madame Saufrany”.

Measured on an RTX 4090 (the only GPU tested), Core i9-14900KF (32 threads) and FFmpeg 8.1 (gyan.dev full build) on 2026-10-07. Audio: LibriVox readings (Gongitsune, CC0; The Gift of the Magi, public domain); reference texts from Aozora Bunko and Project Gutenberg. Japanese: character error rate after converting both texts to kana readings without punctuation or spaces (counting kanji/kana spelling differences too, large-v3-turbo scores 28%). English: word error rate. These are single-speaker readings in a quiet room; conversation, noise or background music give different results. queue was measured at 3 and 20 seconds. Measurements and the start of each transcript are on the data page.