mkdir wespeaker-work && cd $_
uv venv -p 3.12 --seed

uv pip install git+https://github.com/wenet-e2e/wespeaker.git --torch-backend=auto

(snip)
 + wespeaker==0.0.0 (from git+https://github.com/wenet-e2e/wespeaker.git@67f0f4a8d472e6e2203d7baca38daba818af17f3)

uv pip install PyYAML requests

uv run wespeaker -h

usage: wespeaker [-h] [-t {embedding,embedding_kaldi,similarity,diarization,diarization_list}] [-l {chinese,english}]
                 [--campplus] [--eres2net] [--vblinkp] [--vblinkf] [-p PRETRAIN] [--device DEVICE]
                 [--audio_file AUDIO_FILE] [--audio_file2 AUDIO_FILE2] [--wav_scp WAV_SCP]
                 [--resample_rate RESAMPLE_RATE] [--vad] [--output_file OUTPUT_FILE]
                 [--diar_min_duration DIAR_MIN_DURATION] [--diar_window_secs DIAR_WINDOW_SECS]
                 [--diar_period_secs DIAR_PERIOD_SECS] [--diar_frame_shift DIAR_FRAME_SHIFT]
                 [--diar_emb_bs DIAR_EMB_BS] [--diar_subseg_cmn DIAR_SUBSEG_CMN]

options:
  -h, --help            show this help message and exit
  -t {embedding,embedding_kaldi,similarity,diarization,diarization_list}, --task {embedding,embedding_kaldi,similarity,diarization,diarization_list}
                        task type
  -l {chinese,english}, --language {chinese,english}
                        language type
  --campplus            whether to use the damo/speech_campplus_sv_zh-cn_16k-common model
  --eres2net            whether to use the damo/speech_eres2net_sv_zh-cn_16k-common model
  --vblinkp             whether to use the samresnet34 model pretrained on voxblink2
  --vblinkf             whether to use the samresnet34 model pretrained on voxblink2 andfintuned on voxceleb2
  -p PRETRAIN, --pretrain PRETRAIN
                        model directory
  --device DEVICE       device type (most commonly cpu or cuda,but also potentially mps, xpu, xla or meta)and optional
                        device ordinal for the device type.
  --audio_file AUDIO_FILE
                        audio file
  --audio_file2 AUDIO_FILE2
                        audio file2, specifically for similarity task
  --wav_scp WAV_SCP     path to wav.scp, for extract and saving kaldi-stype embeddings
  --resample_rate RESAMPLE_RATE
                        resampling rate
  --vad                 whether to do VAD or not
  --output_file OUTPUT_FILE
                        output file to save speaker embedding or save diarization result
  --diar_min_duration DIAR_MIN_DURATION
                        VAD min duration
  --diar_window_secs DIAR_WINDOW_SECS
                        the window seconds in embedding extraction
  --diar_period_secs DIAR_PERIOD_SECS
                        the shift seconds in embedding extraction
  --diar_frame_shift DIAR_FRAME_SHIFT
                        frame shift in fbank extraction (ms)
  --diar_emb_bs DIAR_EMB_BS
                        batch size for embedding extraction
  --diar_subseg_cmn DIAR_SUBSEG_CMN
                        do cmn after or before fbank sub-segmentation

uv run wespeaker \
    --task embedding \
    --audio_file voice_lunch_jp_1min.wav \
    --output_file embedding.txt

Downloading https://dataset-hub.oss-cn-hangzhou.aliyuncs.com/public-unzip-dataset/wenet/wespeaker_pretrained_models/master/cnceleb_resnet34.tar.gz?Expires=1764275664&OSSAccessKeyId=LTAI5tAoCEDFQFyV5h8unjt8&Signature=Z79oZjXQW2QFdEAHtVVz96gKzAo%3D&response-content-disposition=attachment%3B to /home/kun432/.wespeaker/chinese
cnceleb_resnet34.tar.gz: 100%|█████████████████████████████████████████████████████| 31.1M/31.1M [00:04<00:00, 6.71MB/s]

ImportError: TorchCodec is required for load_with_torchcodec. Please install torchcodec to use this function.

uv pip install torchcodec

(snip)
 + torchcodec==0.8.1

uv run wespeaker \
    --task embedding \
    --audio_file voice_lunch_jp_1min.wav \
    --output_file embedding.txt

Succeed, see embedding.txt

9.884694963693618774e-02
-6.712152808904647827e-02
1.572865545749664307e-01
2.066522091627120972e-01
-3.340280950069427490e-01
(snip)
-1.994613707065582275e-01
-2.906784117221832275e-01
-1.107062846422195435e-01
-5.282661318778991699e-02
-3.054174780845642090e-01

wc -l embedding.txt

256 embedding.txt

git lfs install
git clone https://huggingface.co/Wespeaker/wespeaker-voxceleb-resnet34-LM

uv run wespeaker \
    --pretrain ./wespeaker-voxceleb-resnet34-LM \
    --task embedding \
    --audio_file voice_lunch_jp_1min.wav \
    --output_file embedding2.txt

  File "/work/wespeaker-work/.venv/lib/python3.12/site-packages/wespeaker/cli/speaker.py", line 343, in main
    model = load_model(model_dir=args.pretrain)
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
TypeError: load_model() got an unexpected keyword argument 'model_dir'

        model = load_model(model_name_or_path=args.pretrain)

    raise FileNotFoundError(f"{file} not found in {model_dir}")
FileNotFoundError: avg_model.pt not found in ./wespeaker-voxceleb-resnet34-LM

tree wespeaker-voxceleb-resnet34-LM/

wespeaker-voxceleb-resnet34-LM/
├── README.md
├── avg_model
├── config.yaml
└── voxceleb_resnet34_LM.onnx

0 directories, 4 files

cp wespeaker-voxceleb-resnet34-LM/avg_model wespeaker-voxceleb-resnet34-LM/av
g_model.pt

Succeed, see embedding2.txt

-1.454869508743286133e-01
-1.880786567926406860e-02
2.014929242432117462e-02
2.714627981185913086e-01
8.721305429935455322e-02
(snip)
1.460620611906051636e-01
1.842129677534103394e-01
7.246416062116622925e-03
1.637877374887466431e-01
2.479757070541381836e-01

wc -l embedding2.txt

256 embedding2.txt

wespeaker --task embedding_kaldi --wav_scp wav.scp --output_file /path/to/embedding

uv run wespeaker \
    --pretrain ./wespeaker-voxceleb-resnet34-LM \
    --task similarity \
    --audio_file voice_lunch_jp_1min.wav \
    --audio_file2 my_sample.wav

0.8032963275909424

uv run wespeaker \
    --pretrain ./wespeaker-voxceleb-resnet34-LM \
    --task similarity \
    --audio_file voice_lunch_jp_1min.wav \
    --audio_file2 other_sample.wav

0.532842256128788

uv run wespeaker \
    --pretrain ./wespeaker-voxceleb-resnet34-LM \
    --task similarity \
    --audio_file voice_lunch_jp_1min.wav \
    --audio_file2 voice_lunch_jp_15sec.wav

0.9200428426265717

uv run wespeaker \
    --pretrain ./wespeaker-voxceleb-resnet34-LM \
    --task diarization \
    --audio_file voice_lunch_jp_1min.wav

0.900	6.200	0
6.900	7.900	0
8.100	9.300	0
9.600	10.900	0
11.500	12.500	0
13.900	23.200	0
23.500	26.100	0
26.500	36.600	0
36.700	39.000	0
40.100	40.700	0
40.900	42.800	0

Speaker 1:
[en] What is your best seller here?

Speaker 2:
[en] Our best seller here is cold brew iced coffee and lattes.

Speaker 1:
[en] Okay. And on a day like today, where it's snowing quite a bit, do a lot of people still order iced coffee?

Speaker 2:
[en] Here in Maine, yes.

Speaker 1:
[en] Really?

Speaker 2:
[en] Yes.

ffmpeg -i coffee_shop.mp3 coffee_shop.wav

uv run wespeaker \
    --pretrain ./wespeaker-voxceleb-resnet34-LM \
    --task diarization \
    --audio_file coffee_shop.wav

0.600	2.000	0
2.200	5.400	0
5.600	10.900	0
11.100	16.070	0

uv run wespeaker \
    --pretrain ./wespeaker-voxceleb-resnet34-LM \
    --task diarization \
    --audio_file coffee_shop.wav \
    --vad \
    --diar_min_duration 0.05 \
    --diar_window_secs 0.5 \
    --diar_period_secs 0.25

0.600	1.225	0
1.225	2.000	1
2.200	5.400	1
5.600	8.225	1
8.225	10.475	0
10.475	10.900	1
11.100	12.225	0
12.225	12.475	1
12.475	12.975	0
12.975	16.070	1

import wespeaker

model = wespeaker.load_model("./wespeaker-voxceleb-resnet34-LM")
model.set_device('cuda:0')

embedding = model.extract_embedding('voice_lunch_jp_15sec.wav')
print(type(embedding))
print(embedding.shape)
print(embedding)

<class 'torch.Tensor'>
torch.Size([256])
tensor([-0.0147, -0.0353, -0.2119,  0.2899,  0.0600, -0.0959, -0.3547, -0.3344,
        -0.1263,  0.1023,  0.2098, -0.1215,  0.2277, -0.0788, -0.0625, -0.0447,
         0.2672, -0.0007,  0.1421, -0.0700, -0.1575, -0.0538, -0.1317,  0.0093,
        (snip)
         0.1442, -0.2110, -0.2013, -0.1815, -0.0191,  0.5470,  0.2977, -0.0162,
        -0.0375, -0.3550,  0.0343,  0.1208, -0.1204, -0.0154,  0.1941, -0.2209,
        -0.0923, -0.0446,  0.1504,  0.1633,  0.1414, -0.0021,  0.0977,  0.2701])

import wespeaker

model = wespeaker.load_model("./wespeaker-voxceleb-resnet34-LM")
model.set_device('cuda:0')

similarity1 = model.compute_similarity(
    'voice_lunch_jp_15sec.wav',
    'my_sample.wav'
)
print(similarity1)

similarity2 = model.compute_similarity(
    'voice_lunch_jp_15sec.wav',
    'other_sample.wav'
)
print(similarity2)

0.8291692435741425
0.5287035647779703

import wespeaker

model = wespeaker.load_model("./wespeaker-voxceleb-resnet34-LM")
model.set_device('cuda:0')

diar_result = model.diarize('coffee_shop.wav')
print(diar_result)

[('unk', 0.6, 2.0, 0), ('unk', 2.2, 5.4, 0), ('unk', 5.6, 10.9, 0), ('unk', 11.1, 16.07, 0)]

import wespeaker

model = wespeaker.load_model("./wespeaker-voxceleb-resnet34-LM")
model.set_device('cuda:0')

# VADを有効化
model.set_vad(True)

#VADのパラメータをセット
model.set_diarization_params(  
    min_duration=0.05,
    window_secs=0.5,
    period_secs=0.25,
)

diar_result = model.diarize('coffee_shop.wav')
print(diar_result)

[('unk', 0.6, 2.0, 0), ('unk', 2.2, 5.4, 0), ('unk', 5.6, 6.225, 1), ('unk', 6.225, 10.9, 0), ('unk', 11.1, 13.725, 0), ('unk', 13.725, 13.975, 1), ('unk', 13.975, 16.07, 0)]