The output log in console is not correct with medium model about whisper-standalone-win

[Purfview]

Post a screenshot of what is written when it finishes.

[yuchaoqian]

There is no log, the audio is 17min, and the task always stops in the middle of processing when using the medium model

[Purfview]

And you sure that whisper-faster is not in memory anymore?
Do you run on CUDA? How much VRAM it has?

[yuchaoqian]

Yes it is run on CUDA using 3070. It has 8GB.
During the execution, i saw vram usage is below 4GB in total. Once the task exits, the ressource usage drops immediately.

When i tried the large model, it works just fine with success log

Transcription speed: 9.89 audio seconds/s

Subtitles are written to 'D:\workspace\whisper\fast-whisper\Whisper-Faster' directory.


Operation finished in: 165 seconds

[Purfview]

Strange. Can you share that audio?

[yuchaoqian]

Nah I can not share it.

[Purfview]

In what OS do you run it?
Maybe there would me more info with --verbose=true.
Try --compute_type=int8.

[yuchaoqian]

It is on Windows 10 22H2

2023-07-13 10:55:43.305] [ctranslate2] [thread 30220] [info] CPU: AuthenticAMD (SSE4.1=true, AVX=true, AVX2=true, AVX512=false)
[2023-07-13 10:55:43.305] [ctranslate2] [thread 30220] [info]  - Selected ISA: AVX2
[2023-07-13 10:55:43.305] [ctranslate2] [thread 30220] [info]  - Use Intel MKL: false
[2023-07-13 10:55:43.305] [ctranslate2] [thread 30220] [info]  - SGEMM backend: DNNL (packed: false)
[2023-07-13 10:55:43.305] [ctranslate2] [thread 30220] [info]  - GEMM_S16 backend: none (packed: false)
[2023-07-13 10:55:43.305] [ctranslate2] [thread 30220] [info]  - GEMM_S8 backend: DNNL (packed: false, u8s8 preferred: true)
[2023-07-13 10:55:43.305] [ctranslate2] [thread 30220] [info] GPU #0: NVIDIA GeForce RTX 3070 (CC=8.6)
[2023-07-13 10:55:43.305] [ctranslate2] [thread 30220] [info]  - Allow INT8: true
[2023-07-13 10:55:43.305] [ctranslate2] [thread 30220] [info]  - Allow FP16: true (with Tensor Cores: true)
[2023-07-13 10:55:44.283] [ctranslate2] [thread 30220] [info] Using CUDA allocator: cuda_malloc_async
[2023-07-13 10:55:44.644] [ctranslate2] [thread 30220] [info] Loaded model D:\workspace\whisper\fast-whisper\Whisper-Faster\_models\faster-whisper-medium on device cuda:0
[2023-07-13 10:55:44.644] [ctranslate2] [thread 30220] [info]  - Binary version: 6
[2023-07-13 10:55:44.644] [ctranslate2] [thread 30220] [info]  - Model specification revision: 3
[2023-07-13 10:55:44.644] [ctranslate2] [thread 30220] [info]  - Selected compute type: float16

Standalone Faster-Whisper r134.6 running on: CUDA

Number of visible GPU devices: 1

Supported compute types by GPU: {'float16', 'int8', 'int8_float16', 'float32'}


Model loaded in: 1.41 seconds

Processing audio with duration 17:38.773

VAD filter removed 00:31.889 of audio
VAD filter kept the following audio segments: [00:00.000 -> 17:06.884]

Audio processing finished in: 5.02 seconds

Processing segment at 00:00.000

# ....same Processing segment log here so I do not paste them here
# and we have some compression ratio log here
Compression ratio threshold is not met with temperature 0.0 (8.540984 > 2.400000)
Compression ratio threshold is not met with temperature 0.2 (8.540984 > 2.400000)
Compression ratio threshold is not met with temperature 0.4 (8.540984 > 2.400000)
Compression ratio threshold is not met with temperature 0.6 (7.133333 > 2.400000)
Compression ratio threshold is not met with temperature 0.8 (7.444444 > 2.400000)

Processing segment at 16:42.470

And it exits without any error. Just ring a "dou" song from the speaker.

[Purfview]

But it shows that it's working -> "Processing segment at 16:42.470".
If you hear beep then that indicates that it's successfully finished, and there should be written "Operation finished".

[yuchaoqian]

I think the output log is not correct.
There is no log be written "Operation finished" when using medium size model even it is successfully finished.
I could find the text from 16:42.470 to the end in the generated srt file after the beep, but the last line of log is "Processing segment at 16:42.470".

And the situation is even worse when not using "--verbose"
When it beep, the logged lines never reach the end, but the srt file is correct.

[10:21.460 --> 10:22.460]

[Purfview]

Created .srt and "beep" indicates that a task is successfully finished.
Just that the log output to console for some reason doesn't show up for you.

Check if r136.7 corrected this issue.