--diarize flag no longer works with stereo input in latest release

[varlog11]

Environment:
whisper.cpp version: latest (1.7.5)
Platform: Debian 12, x86_64
Build flags: -D WHISPER_FFMPEG=yes -D GGML_CUDA=no
Test file: 2-channel stereo WAV with two clean channels

In version 1.7.4, the --diarize option correctly performed voice diarization on a stereo input file (splitting tracks by speaker/channel). After upgrading to the latest release, the same command no longer produces separate speaker outputs.

[danbev]

Could you provide a sample audio file that I can use to reproduce this, and also what model are you using?

[Zeugma4242]

Regression example of --diarize stereo splitting between v1.7.4 and v1.7.5 (French audio)

I’m seeing the same issue on macOS (Apple M1, Metal backend) when transcribing a 2-channel WAV in French: in v1.7.5 , the transcription is garbage text, while in v1.7.4 the transcription is correct and of course speakers are also diarized correctly.

---

Environment

• Version: v1.7.5 or v.1.7.4 tags ; built with make -j8 WHISPER_METAL=1
• Model: ggml-medium.bin
• Test file: test.wav (2-channel, 16 kHz)

Command line

whisper-cli \
  -m /path/to/ggml-medium.bin \
  --language fr \
  --beam-size 5 \
  --best-of 5 \
  --max-context 20 \
  --output-json-full \
  --output-txt \
  --print-progress \
  --diarize \
  -f /path/to/test.wav \
  -of /path/to/output

---

v1.7.5 output (all segments end up as non sense garbage - usually references of sounds of legal mentions like this actual example)

[00:00:00.000 --> 00:00:03.000]  (speaker 1) Sous-titres réalisés para la communauté d'Amara.org
…

---

v1.7.4 output (correct text and speakers added)

[00:00:00.000 --> 00:00:01.340]  (speaker 1) [sonnerie]
[00:00:01.340 --> 00:00:05.600]  (speaker 1) [silence]
[00:00:05.600 --> 00:00:08.880]  (speaker 1) — Salut, jeune homme. — C'est vrai que je te dérange ?
[00:00:08.880 --> 00:00:10.480]  (speaker 1) — Ah pas du tout, pas du tout, pas du tout.
[00:00:10.480 --> 00:00:14.020]  (speaker 1) J'étais en train de préparer un courrier.
…

---

Notes

• Not all stereo WAVs exhibit the issue, but it’s 100% reproducible on my test.wav.
• Same behavior on some English audio with --language en in parameters
• In v1.7.4 it always works as expected.

test.wav to reproduce is attached to this post. Let me know if you need additional logs or tests to help fix the issue.

test.wav.zip

[danbev]

@Zeugma4242 Thanks for the details on this and the test file, this is very helpful. I'm looking into this now.