Just words of encouragement: the quality of diarization is very high

[Manamama]

An end-user mostly, but see e.g. numpy/numpy#27698 (comment), serengil/deepface#1614 that sometimes I delve into code.

• I have just advised all and sundry to use yours: https://github.com/OpenMOSS/MOSS-Transcribe-Diarize.git
• It installs (and runs) on Termux! (after one hour of upgrading av and transformers and the like: expected : Successfully installed audioop-lts-0.2.2 audioread-3.1.0 av-18.0.0 decorator-5.3.1 fastapi-0.139.2 joblib-1.5.3 librosa-0.11.0 moss-transcribe-diarize-0.1.0 pooch-1.9.0 python-multipart-0.0.32 scikit-learn-1.9.0 soundfile-0.14.0 soxr-1.1.0 standard-aifc-3.13.0 standard-chunk-3.13.0 standard-sunau-3.13.0 starlette-1.3.1 threadpoolctl-3.6.0 tokenizers-0.22.2 transformers-5.14.1 uvicorn-0.51.0 , on a humble Linux localhost 4.14.186+ #1 SMP PREEMPT Thu Mar 17 16:28:22 CST 2022 aarch64 Android "

~/Downloads/GitHub $ mtd-subtitle-web --model OpenMOSS-Team/MOSS-Transcribe-Diarize 
WARNING: linker: Warning: "/data/data/com.termux/files/usr/lib/python3.14/site-packages/optree.libs/libc++_shared-d523468d.so" unused DT entry: unknown processor-specific (type 0x70000001 arg 0x0) (ignoring)
INFO:     Started server process [9017]
INFO:     Waiting for application startup.
INFO:     Application startup complete.
INFO:     Uvicorn running on http://127.0.0.1:7860 (Press CTRL+C to quit)
INFO:     127.0.0.1:53816 - "GET / HTTP/1.1" 200 OK
INFO:     127.0.0.1:53816 - "GET /api/runtime HTTP/1.1" 200 OK
INFO:     127.0.0.1:53818 - "GET /api/jobs HTTP/1.1" 200 OK
INFO:     127.0.0.1:53818 - "GET /favicon.svg HTTP/1.1" 200 OK
INFO:     127.0.0.1:53838 - "POST /api/jobs HTTP/1.1" 200 OK
INFO:     127.0.0.1:53838 - "GET /api/jobs HTTP/1.1" 200 OK
INFO:     127.0.0.1:53838 - "GET /api/jobs/6ba85a15c0a7 HTTP/1.1" 200 OK
INFO:     127.0.0.1:53838 - "GET /api/jobs HTTP/1.1" 200 OK
Loading weights:   9%|██████████████▍                                                                                                                                           | 64/683 [00:00<00:02, 250.07it/s]INFO:     127.0.0.1:53838 - "GET /api/jobs HTTP/1.1" 200 OK
Loading weights:  82%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████                           | 563/683 [00:01<00:00, 416.18it/s]INFO:     127.0.0.1:53838 - "GET /api/jobs HTTP/1.1" 200 OK
Loading weights: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 683/683 [00:01<00:00, 352.47it/s]
INFO:     127.0.0.1:53838 - "GET /api/jobs HTTP/1.1" 200 OK
INFO:     127.0.0.1:53838 - "GET /api/jobs HTTP/1.1" 200 OK
[transformers] `MossTranscribeDiarizeProcessor` defines `feature_extractor_class = 'AutoFeatureExtractor'`, which is deprecated. Register the correct mapping in `AutoFeatureExtractor` instead.
INFO:     127.0.0.1:53838 - "GET /api/jobs HTTP/1.1" 200 OK
INFO:     127.0.0.1:53838 - "GET /api/jobs HTTP/1.1" 200 OK
INFO:     127.0.0.1:53838 - "GET /api/jobs HTTP/1.1" 200 OK
INFO:     127.0.0.1:53838 - "GET /api/jobs HTTP/1.1" 200 OK

• The GUI (after the transcription be done) is excellent too: I coded something with whisperx and PANN and plotly dynamic graphs years ago, while yours is clean and nifty.
• The GUI helps one learn Chinese ;).
• The diarization quality is excellent. Many files that had at best mangled most speakers to one, or the opposite: multiplied them as rabbits, via whisperx like pipelines even, now properly separate these to the right buckets. (Yes, I must read your paper how so, in detail, one day...)

1
00:00:05,330 --> 00:00:07,330
S01: Sognia's head, head!

2
00:00:08,060 --> 00:00:08,830
S01: Roder!

3
00:00:09,100 --> 00:00:11,100
S01: Sognia's head!

4
00:00:12,930 --> 00:00:13,660
S01: Roder!

5
00:00:13,700 --> 00:00:15,700
S01: Sognia's head, Sognia!

6
00:01:14,030 --> 00:01:18,830
S02: The man with the green jacket, the one who killed Zaper, find out what he knows.

7
00:01:38,760 --> 00:01:39,430
S03: Hi!

8
00:01:41,300 --> 00:01:42,830
S03: Hello, hello, A-Boy.

9
00:01:43,760 --> 00:01:44,530
S03: You alone?

10
00:01:46,000 --> 00:01:47,330
S04: Jätkaan blade raan.

11
00:01:48,230 --> 00:01:49,030
S04: Mitä se on?

12
00:01:50,260 --> 00:01:51,330
S04: Mennä, se on pitu vaan.

13
00:01:51,330 --> 00:01:51,830
S03: It's okay.

14
00:01:52,530 --> 00:01:53,130
S03: It's fine.

15
00:02:00,130 --> 00:02:01,530
S03: Wanna buy a lady a cigarette?

16
00:02:08,830 --> 00:02:10,260
S03: Oh, you don't even smile.

17
00:02:13,260 --> 00:02:14,630
S05: Didn't you hear your friends?

18
00:02:15,630 --> 00:02:16,730
S05: Don't you know what I am?

19
00:02:17,730 --> 00:02:18,430
S03: Yeah.

20
00:02:19,560 --> 00:02:20,930
S03: Guy eating rice.

21
00:02:22,760 --> 00:02:23,430
S03: What's that?

22
00:02:25,430 --> 00:02:26,130
S05: It's a tree.

23
00:02:27,230 --> 00:02:27,730
S03: Oh.

24
00:02:28,260 --> 00:02:29,730
S03: Never seen a tree before.

25
00:02:31,930 --> 00:02:32,630
S03: It's pretty.

26
00:02:35,430 --> 00:02:36,230
S05: It's dead.

27
00:02:38,630 --> 00:02:40,130
S03: Now, who keeps...

28
00:02:40,930 --> 00:02:42,430
S03: A dead tree?

29
00:02:50,360 --> 00:02:51,830
S03: You're not gonna kill me, are you?

30
00:02:52,160 --> 00:02:52,930
S05: Depends.

31
00:02:53,930 --> 00:02:54,930
S05: What's your model number?

32
00:02:55,760 --> 00:02:57,730
S03: Why don't you look into my eye and find out?

33
00:03:00,860 --> 00:03:02,130
S03: Oh...

34
00:03:02,930 --> 00:03:04,330
S03: You don't like girl girls.

35
00:03:08,230 --> 00:03:08,730
S03: Well...

36
00:03:09,430 --> 00:03:10,260
S03: I'm always here.

from a famous movie, doable without any optimization (scaling down, lower sized models, switches) on:

• 再次恭喜 !

[zimoCCC]

Thank you so, so, so, so, so much; we will continue to work hard.

As for “The GUI helps one learn Chinese ;),” that was an oversight on my part. I’ll switch the GUI to English as soon as possible.

[zimoCCC]

The subtitle Web UI now supports both Simplified Chinese and English.

[rajeevbarde]

big time !!