PR #22553 opened by WyattBlue URL: https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/22553 Patch URL: https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/22553.patch
This prevents `max_len` splitting via tokens, which splits words like "don't" and proper nouns inappropriately. >From d1589c69b25810d828d0bec969c5234e6cc8b050 Mon Sep 17 00:00:00 2001 From: WyattBlue <[email protected]> Date: Thu, 19 Mar 2026 18:01:45 -0400 Subject: [PATCH] avfilter/af_whisper.c: Set split_on_word This prevents `max_len` splitting via tokens, which splits words like "don't" and proper nouns inappropriately. --- doc/filters.texi | 4 ++-- libavfilter/af_whisper.c | 1 + 2 files changed, 3 insertions(+), 2 deletions(-) diff --git a/doc/filters.texi b/doc/filters.texi index 569ff516d4..3b9fd893e0 100644 --- a/doc/filters.texi +++ b/doc/filters.texi @@ -7775,8 +7775,8 @@ Default value: @code{"text"} @item max_len Maximum segment length in characters. When set to a value greater than 0, -transcription segments will be split to not exceed this length. This is useful -for generating subtitles with shorter lines. +transcription segments will be split by word to not exceed this length. This is +useful for generating subtitles with shorter lines. Default value: @code{"0"} @item vad_model diff --git a/libavfilter/af_whisper.c b/libavfilter/af_whisper.c index 299a8bca7a..e5723b30cf 100644 --- a/libavfilter/af_whisper.c +++ b/libavfilter/af_whisper.c @@ -207,6 +207,7 @@ static void run_transcription(AVFilterContext *ctx, AVFrame *frame, int samples) params.print_timestamps = 0; params.max_len = wctx->max_len; params.token_timestamps = (wctx->max_len > 0); + params.split_on_word = (wctx->max_len > 0); if (whisper_full(wctx->ctx_wsp, params, wctx->audio_buffer, samples) != 0) { av_log(ctx, AV_LOG_ERROR, "Failed to process audio with whisper.cpp\n"); -- 2.52.0 _______________________________________________ ffmpeg-devel mailing list -- [email protected] To unsubscribe send an email to [email protected]
