Initial commit

examples/libritts/cosyvoice/tts_text.json

+{
+  "1089_134686_000002_000000": [
+    "hello, my name is Jack. What is your name?"
+  ]
+}
\ No newline at end of file

examples/magicdata-read/cosyvoice/conf/cosyvoice.fromscratch.yaml

+# set random seed, so that you may reproduce your result.
+__set_seed1: !apply:random.seed [1986]
+__set_seed2: !apply:numpy.random.seed [1986]
+__set_seed3: !apply:torch.manual_seed [1986]
+__set_seed4: !apply:torch.cuda.manual_seed_all [1986]
+
+# fixed params
+sample_rate: 22050
+text_encoder_input_size: 512
+llm_input_size: 1024
+llm_output_size: 1024
+spk_embed_dim: 192
+
+# model params
+# for all class/function included in this repo, we use !<name> or !<new> for intialization, so that user may find all corresponding class/function according to one single yaml.
+# for system/third_party class/function, we do not require this.
+llm: !new:cosyvoice.llm.llm.TransformerLM
+    text_encoder_input_size: !ref <text_encoder_input_size>
+    llm_input_size: !ref <llm_input_size>
+    llm_output_size: !ref <llm_output_size>
+    text_token_size: 51866
+    speech_token_size: 4096
+    length_normalized_loss: True
+    lsm_weight: 0
+    spk_embed_dim: !ref <spk_embed_dim>
+    text_encoder: !new:cosyvoice.transformer.encoder.ConformerEncoder
+        input_size: !ref <text_encoder_input_size>
+        output_size: 1024
+        attention_heads: 8
+        linear_units: 2048
+        num_blocks: 3
+        dropout_rate: 0.1
+        positional_dropout_rate: 0.1
+        attention_dropout_rate: 0
+        normalize_before: True
+        input_layer: 'linear'
+        pos_enc_layer_type: 'rel_pos_espnet'
+        selfattention_layer_type: 'rel_selfattn'
+        use_cnn_module: False
+        macaron_style: False
+        use_dynamic_chunk: False
+        use_dynamic_left_chunk: False
+        static_chunk_size: 1
+    llm: !new:cosyvoice.transformer.encoder.TransformerEncoder
+        input_size: !ref <llm_input_size>
+        output_size: !ref <llm_output_size>
+        attention_heads: 8
+        linear_units: 2048
+        num_blocks: 7
+        dropout_rate: 0.1
+        positional_dropout_rate: 0.1
+        attention_dropout_rate: 0
+        input_layer: 'linear_legacy'
+        pos_enc_layer_type: 'rel_pos_espnet'
+        selfattention_layer_type: 'rel_selfattn'
+        static_chunk_size: 1
+
+flow: !new:cosyvoice.flow.flow.MaskedDiffWithXvec
+    input_size: 512
+    output_size: 80
+    spk_embed_dim: !ref <spk_embed_dim>
+    output_type: 'mel'
+    vocab_size: 4096
+    input_frame_rate: 50
+    only_mask_loss: True
+    encoder: !new:cosyvoice.transformer.encoder.ConformerEncoder
+        output_size: 512
+        attention_heads: 4
+        linear_units: 1024
+        num_blocks: 3
+        dropout_rate: 0.1
+        positional_dropout_rate: 0.1
+        attention_dropout_rate: 0.1
+        normalize_before: True
+        input_layer: 'linear'
+        pos_enc_layer_type: 'rel_pos_espnet'
+        selfattention_layer_type: 'rel_selfattn'
+        input_size: 512
+        use_cnn_module: False
+        macaron_style: False
+    length_regulator: !new:cosyvoice.flow.length_regulator.InterpolateRegulator
+        channels: 80
+        sampling_ratios: [1, 1, 1, 1]
+    decoder: !new:cosyvoice.flow.flow_matching.ConditionalCFM
+        in_channels: 240
+        n_spks: 1
+        spk_emb_dim: 80
+        cfm_params: !new:omegaconf.DictConfig
+            content:
+                sigma_min: 1e-06
+                solver: 'euler'
+                t_scheduler: 'cosine'
+                training_cfg_rate: 0.2
+                inference_cfg_rate: 0.7
+                reg_loss_type: 'l1'
+        estimator: !new:cosyvoice.flow.decoder.ConditionalDecoder
+            in_channels: 320
+            out_channels: 80
+            channels: [256, 256]
+            dropout: 0
+            attention_head_dim: 64
+            n_blocks: 4
+            num_mid_blocks: 8
+            num_heads: 8
+            act_fn: 'gelu'
+
+hift: !new:cosyvoice.hifigan.generator.HiFTGenerator
+    in_channels: 80
+    base_channels: 512
+    nb_harmonics: 8
+    sampling_rate: !ref <sample_rate>
+    nsf_alpha: 0.1
+    nsf_sigma: 0.003
+    nsf_voiced_threshold: 10
+    upsample_rates: [8, 8]
+    upsample_kernel_sizes: [16, 16]
+    istft_params:
+        n_fft: 16
+        hop_len: 4
+    resblock_kernel_sizes: [3, 7, 11]
+    resblock_dilation_sizes: [[1, 3, 5], [1, 3, 5], [1, 3, 5]]
+    source_resblock_kernel_sizes: [7, 11]
+    source_resblock_dilation_sizes: [[1, 3, 5], [1, 3, 5]]
+    lrelu_slope: 0.1
+    audio_limit: 0.99
+    f0_predictor: !new:cosyvoice.hifigan.f0_predictor.ConvRNNF0Predictor
+        num_class: 1
+        in_channels: 80
+        cond_channels: 512
+
+# processor functions
+parquet_opener: !name:cosyvoice.dataset.processor.parquet_opener
+get_tokenizer: !name:whisper.tokenizer.get_tokenizer
+    multilingual: True
+    num_languages: 100
+    language: 'en'
+    task: 'transcribe'
+allowed_special: 'all'
+tokenize: !name:cosyvoice.dataset.processor.tokenize
+    get_tokenizer: !ref <get_tokenizer>
+    allowed_special: !ref <allowed_special>
+filter: !name:cosyvoice.dataset.processor.filter
+    max_length: 40960
+    min_length: 0
+    token_max_length: 200
+    token_min_length: 1
+resample: !name:cosyvoice.dataset.processor.resample
+    resample_rate: !ref <sample_rate>
+feat_extractor: !name:matcha.utils.audio.mel_spectrogram
+    n_fft: 1024
+    num_mels: 80
+    sampling_rate: !ref <sample_rate>
+    hop_size: 256
+    win_size: 1024
+    fmin: 0
+    fmax: 8000
+    center: False
+compute_fbank: !name:cosyvoice.dataset.processor.compute_fbank
+    feat_extractor: !ref <feat_extractor>
+parse_embedding: !name:cosyvoice.dataset.processor.parse_embedding
+    normalize: True
+shuffle: !name:cosyvoice.dataset.processor.shuffle
+    shuffle_size: 1000
+sort: !name:cosyvoice.dataset.processor.sort
+    sort_size: 500  # sort_size should be less than shuffle_size
+batch: !name:cosyvoice.dataset.processor.batch
+    batch_type: 'dynamic'
+    max_frames_in_batch: 12000
+padding: !name:cosyvoice.dataset.processor.padding
+    use_spk_embedding: False # change to True during sft
+
+# dataset processor pipeline
+data_pipeline: [
+    !ref <parquet_opener>,
+    !ref <tokenize>,
+    !ref <filter>,
+    !ref <resample>,
+    !ref <compute_fbank>,
+    !ref <parse_embedding>,
+    !ref <shuffle>,
+    !ref <sort>,
+    !ref <batch>,
+    !ref <padding>,
+]
+
+# train conf
+train_conf:
+    optim: adam
+    optim_conf:
+        lr: 0.002 # change to 0.001 if you want to train flow from scratch
+    scheduler: warmuplr
+    scheduler_conf:
+        warmup_steps: 25000
+    max_epoch: 200
+    grad_clip: 5
+    accum_grad: 2
+    log_interval: 100
+    save_per_step: -1
\ No newline at end of file

examples/magicdata-read/cosyvoice/conf/cosyvoice.yaml

+# set random seed, so that you may reproduce your result.
+__set_seed1: !apply:random.seed [1986]
+__set_seed2: !apply:numpy.random.seed [1986]
+__set_seed3: !apply:torch.manual_seed [1986]
+__set_seed4: !apply:torch.cuda.manual_seed_all [1986]
+
+# fixed params
+sample_rate: 22050
+text_encoder_input_size: 512
+llm_input_size: 1024
+llm_output_size: 1024
+spk_embed_dim: 192
+
+# model params
+# for all class/function included in this repo, we use !<name> or !<new> for intialization, so that user may find all corresponding class/function according to one single yaml.
+# for system/third_party class/function, we do not require this.
+llm: !new:cosyvoice.llm.llm.TransformerLM
+    text_encoder_input_size: !ref <text_encoder_input_size>
+    llm_input_size: !ref <llm_input_size>
+    llm_output_size: !ref <llm_output_size>
+    text_token_size: 51866
+    speech_token_size: 4096
+    length_normalized_loss: True
+    lsm_weight: 0
+    spk_embed_dim: !ref <spk_embed_dim>
+    text_encoder: !new:cosyvoice.transformer.encoder.ConformerEncoder
+        input_size: !ref <text_encoder_input_size>
+        output_size: 1024
+        attention_heads: 16
+        linear_units: 4096
+        num_blocks: 6
+        dropout_rate: 0.1
+        positional_dropout_rate: 0.1
+        attention_dropout_rate: 0
+        normalize_before: True
+        input_layer: 'linear'
+        pos_enc_layer_type: 'rel_pos_espnet'
+        selfattention_layer_type: 'rel_selfattn'
+        use_cnn_module: False
+        macaron_style: False
+        use_dynamic_chunk: False
+        use_dynamic_left_chunk: False
+        static_chunk_size: 1
+    llm: !new:cosyvoice.transformer.encoder.TransformerEncoder
+        input_size: !ref <llm_input_size>
+        output_size: !ref <llm_output_size>
+        attention_heads: 16
+        linear_units: 4096
+        num_blocks: 14
+        dropout_rate: 0.1
+        positional_dropout_rate: 0.1
+        attention_dropout_rate: 0
+        input_layer: 'linear_legacy'
+        pos_enc_layer_type: 'rel_pos_espnet'
+        selfattention_layer_type: 'rel_selfattn'
+        static_chunk_size: 1
+
+flow: !new:cosyvoice.flow.flow.MaskedDiffWithXvec
+    input_size: 512
+    output_size: 80
+    spk_embed_dim: !ref <spk_embed_dim>
+    output_type: 'mel'
+    vocab_size: 4096
+    input_frame_rate: 50
+    only_mask_loss: True
+    encoder: !new:cosyvoice.transformer.encoder.ConformerEncoder
+        output_size: 512
+        attention_heads: 8
+        linear_units: 2048
+        num_blocks: 6
+        dropout_rate: 0.1
+        positional_dropout_rate: 0.1
+        attention_dropout_rate: 0.1
+        normalize_before: True
+        input_layer: 'linear'
+        pos_enc_layer_type: 'rel_pos_espnet'
+        selfattention_layer_type: 'rel_selfattn'
+        input_size: 512
+        use_cnn_module: False
+        macaron_style: False
+    length_regulator: !new:cosyvoice.flow.length_regulator.InterpolateRegulator
+        channels: 80
+        sampling_ratios: [1, 1, 1, 1]
+    decoder: !new:cosyvoice.flow.flow_matching.ConditionalCFM
+        in_channels: 240
+        n_spks: 1
+        spk_emb_dim: 80
+        cfm_params: !new:omegaconf.DictConfig
+            content:
+                sigma_min: 1e-06
+                solver: 'euler'
+                t_scheduler: 'cosine'
+                training_cfg_rate: 0.2
+                inference_cfg_rate: 0.7
+                reg_loss_type: 'l1'
+        estimator: !new:cosyvoice.flow.decoder.ConditionalDecoder
+            in_channels: 320
+            out_channels: 80
+            channels: [256, 256]
+            dropout: 0
+            attention_head_dim: 64
+            n_blocks: 4
+            num_mid_blocks: 12
+            num_heads: 8
+            act_fn: 'gelu'
+
+hift: !new:cosyvoice.hifigan.generator.HiFTGenerator
+    in_channels: 80
+    base_channels: 512
+    nb_harmonics: 8
+    sampling_rate: !ref <sample_rate>
+    nsf_alpha: 0.1
+    nsf_sigma: 0.003
+    nsf_voiced_threshold: 10
+    upsample_rates: [8, 8]
+    upsample_kernel_sizes: [16, 16]
+    istft_params:
+        n_fft: 16
+        hop_len: 4
+    resblock_kernel_sizes: [3, 7, 11]
+    resblock_dilation_sizes: [[1, 3, 5], [1, 3, 5], [1, 3, 5]]
+    source_resblock_kernel_sizes: [7, 11]
+    source_resblock_dilation_sizes: [[1, 3, 5], [1, 3, 5]]
+    lrelu_slope: 0.1
+    audio_limit: 0.99
+    f0_predictor: !new:cosyvoice.hifigan.f0_predictor.ConvRNNF0Predictor
+        num_class: 1
+        in_channels: 80
+        cond_channels: 512
+
+# processor functions
+parquet_opener: !name:cosyvoice.dataset.processor.parquet_opener
+get_tokenizer: !name:whisper.tokenizer.get_tokenizer
+    multilingual: True
+    num_languages: 100
+    language: 'en'
+    task: 'transcribe'
+allowed_special: 'all'
+tokenize: !name:cosyvoice.dataset.processor.tokenize
+    get_tokenizer: !ref <get_tokenizer>
+    allowed_special: !ref <allowed_special>
+filter: !name:cosyvoice.dataset.processor.filter
+    max_length: 40960
+    min_length: 0
+    token_max_length: 200
+    token_min_length: 1
+resample: !name:cosyvoice.dataset.processor.resample
+    resample_rate: !ref <sample_rate>
+feat_extractor: !name:matcha.utils.audio.mel_spectrogram
+    n_fft: 1024
+    num_mels: 80
+    sampling_rate: !ref <sample_rate>
+    hop_size: 256
+    win_size: 1024
+    fmin: 0
+    fmax: 8000
+    center: False
+compute_fbank: !name:cosyvoice.dataset.processor.compute_fbank
+    feat_extractor: !ref <feat_extractor>
+parse_embedding: !name:cosyvoice.dataset.processor.parse_embedding
+    normalize: True
+shuffle: !name:cosyvoice.dataset.processor.shuffle
+    shuffle_size: 1000
+sort: !name:cosyvoice.dataset.processor.sort
+    sort_size: 500  # sort_size should be less than shuffle_size
+batch: !name:cosyvoice.dataset.processor.batch
+    batch_type: 'dynamic'
+    max_frames_in_batch: 2000
+padding: !name:cosyvoice.dataset.processor.padding
+    use_spk_embedding: False # change to True during sft
+
+# dataset processor pipeline
+data_pipeline: [
+    !ref <parquet_opener>,
+    !ref <tokenize>,
+    !ref <filter>,
+    !ref <resample>,
+    !ref <compute_fbank>,
+    !ref <parse_embedding>,
+    !ref <shuffle>,
+    !ref <sort>,
+    !ref <batch>,
+    !ref <padding>,
+]
+
+# train conf
+train_conf:
+    optim: adam
+    optim_conf:
+        lr: 0.001 # change to 1e-5 during sft
+    scheduler: warmuplr # change to constantlr during sft
+    scheduler_conf:
+        warmup_steps: 2500
+    max_epoch: 200
+    grad_clip: 5
+    accum_grad: 2
+    log_interval: 100
+    save_per_step: -1
\ No newline at end of file

examples/magicdata-read/cosyvoice/conf/ds_stage2.json

+{
+  "train_micro_batch_size_per_gpu": 1,
+  "gradient_accumulation_steps": 1,
+  "steps_per_print": 100,
+  "gradient_clipping": 5,
+  "fp16": {
+    "enabled": false,
+    "auto_cast": false,
+    "loss_scale": 0,
+    "initial_scale_power": 16,
+    "loss_scale_window": 256,
+    "hysteresis": 2,
+    "consecutive_hysteresis": false,
+    "min_loss_scale": 1
+  },
+  "bf16": {
+    "enabled": false
+  },
+  "zero_force_ds_cpu_optimizer": false,
+  "zero_optimization": {
+    "stage": 2,
+    "offload_optimizer": {
+      "device": "none",
+      "pin_memory": true
+    },
+    "allgather_partitions": true,
+    "allgather_bucket_size": 5e8,
+    "overlap_comm": false,
+    "reduce_scatter": true,
+    "reduce_bucket_size": 5e8,
+    "contiguous_gradients" : true
+  },
+  "optimizer": {
+    "type": "AdamW",
+    "params": {
+        "lr": 0.001,
+        "weight_decay": 0.0001,
+        "torch_adam": true,
+        "adam_w_mode": true
+    }
+  }
+}
\ No newline at end of file

examples/magicdata-read/cosyvoice/cosyvoice

+../../../cosyvoice
\ No newline at end of file

examples/magicdata-read/cosyvoice/local/download_and_untar.sh

+#!/bin/bash
+
+# Copyright   2014  Johns Hopkins University (author: Daniel Povey)
+# Apache 2.0
+
+remove_archive=false
+
+if [ "$1" == --remove-archive ]; then
+  remove_archive=true
+  shift
+fi
+
+if [ $# -ne 3 ]; then
+  echo "Usage: $0 [--remove-archive] <data-base> <url-base> <corpus-part>"
+  echo "e.g.: $0 /export/a15/vpanayotov/data www.openslr.org/resources/11 dev-clean"
+  echo "With --remove-archive it will remove the archive after successfully un-tarring it."
+  echo "<corpus-part> can be one of: dev-clean, test-clean, dev-other, test-other,"
+  echo "          train-clean-100, train-clean-360, train-other-500."
+  exit 1
+fi
+
+data=$1
+url=$2
+part=$3
+
+if [ ! -d "$data" ]; then
+  echo "$0: no such directory $data"
+  exit 1
+fi
+
+part_ok=false
+list="dev_set test_set train_set"
+for x in $list; do
+  if [ "$part" == $x ]; then part_ok=true; fi
+done
+if ! $part_ok; then
+  echo "$0: expected <corpus-part> to be one of $list, but got '$part'"
+  exit 1
+fi
+
+if [ -z "$url" ]; then
+  echo "$0: empty URL base."
+  exit 1
+fi
+
+if [ -f $data/.$part.complete ]; then
+  echo "$0: data part $part was already successfully extracted, nothing to do."
+  exit 0
+fi
+
+
+# sizes of the archive files in bytes.  This is some older versions.
+sizes_old="1035537823 2201936013 52627842921"
+# sizes_new is the archive file sizes of the final release.  Some of these sizes are of
+# things we probably won't download.
+sizes_new="3886385"
+
+if [ -f $data/$part.tar.gz ]; then
+  size=$(/bin/ls -l $data/$part.tar.gz | awk '{print $5}')
+  size_ok=false
+  for s in $sizes_old $sizes_new; do if [ $s == $size ]; then size_ok=true; fi; done
+  if ! $size_ok; then
+    echo "$0: removing existing file $data/$part.tar.gz because its size in bytes $size"
+    echo "does not equal the size of one of the archives."
+    rm $data/$part.tar.gz
+  else
+    echo "$data/$part.tar.gz exists and appears to be complete."
+  fi
+fi
+
+if [ ! -f $data/$part.tar.gz ]; then
+  if ! which wget >/dev/null; then
+    echo "$0: wget is not installed."
+    exit 1
+  fi
+  full_url=$url/$part.tar.gz
+  echo "$0: downloading data from $full_url.  This may take some time, please be patient."
+
+  if ! wget -P $data --no-check-certificate $full_url; then
+    echo "$0: error executing wget $full_url"
+    exit 1
+  fi
+fi
+
+if ! tar -C $data -xvzf $data/$part.tar.gz; then
+  echo "$0: error un-tarring archive $data/$part.tar.gz"
+  exit 1
+fi
+
+touch $data/.$part.complete
+
+echo "$0: Successfully downloaded and un-tarred $data/$part.tar.gz"
+
+if $remove_archive; then
+  echo "$0: removing $data/$part.tar.gz file since --remove-archive option was supplied."
+  rm $data/$part.tar.gz
+fi

examples/magicdata-read/cosyvoice/local/prepare_data.py

+import argparse
+import logging
+import os
+from tqdm import tqdm
+
+
+logger = logging.getLogger()
+
+def main():
+    utt2wav, utt2text, utt2spk, spk2utt = {}, {}, {}, {}
+    with open(os.path.join(args.src_dir, "TRANS.txt"), "r") as f:
+        lines = f.readlines()[1:]
+        lines = [l.split('\t') for l in lines]
+    for wav, spk, content in tqdm(lines):
+        wav, spk, content = wav.strip(), spk.strip(), content.strip()
+        content = content.replace('[FIL]', '')
+        content = content.replace('[SPK]', '')
+        wav = os.path.join(args.src_dir, spk, wav)
+        if not os.path.exists(wav):
+            continue
+        utt = os.path.basename(wav).replace('.wav', '')
+        utt2wav[utt] = wav
+        utt2text[utt] = content
+        utt2spk[utt] = spk
+        if spk not in spk2utt:
+            spk2utt[spk] = []
+        spk2utt[spk].append(utt)
+
+    with open('{}/wav.scp'.format(args.des_dir), 'w') as f:
+        for k, v in utt2wav.items():
+            f.write('{} {}\n'.format(k, v))
+    with open('{}/text'.format(args.des_dir), 'w') as f:
+        for k, v in utt2text.items():
+            f.write('{} {}\n'.format(k, v))
+    with open('{}/utt2spk'.format(args.des_dir), 'w') as f:
+        for k, v in utt2spk.items():
+            f.write('{} {}\n'.format(k, v))
+    with open('{}/spk2utt'.format(args.des_dir), 'w') as f:
+        for k, v in spk2utt.items():
+            f.write('{} {}\n'.format(k, ' '.join(v)))
+    return
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--src_dir',
+                        type=str)
+    parser.add_argument('--des_dir',
+                        type=str)
+    args = parser.parse_args()
+    main()

examples/magicdata-read/cosyvoice/path.sh

+# NOTE(kan-bayashi): Use UTF-8 in Python to avoid UnicodeDecodeError when LC_ALL=C
+export PYTHONIOENCODING=UTF-8
+export PYTHONPATH=../../../:../../../third_party/Matcha-TTS:$PYTHONPATH

examples/magicdata-read/cosyvoice/run.sh

+#!/bin/bash
+# Copyright 2024 Alibaba Inc. All Rights Reserved.
+. ./path.sh || exit 1;
+
+stage=-1
+stop_stage=3
+
+data_url=www.openslr.org/resources/68
+data_dir=/mnt/hengwu.zty/data/tts/openslr/magicdata-read
+pretrained_model_dir=../../../pretrained_models/CosyVoice-300M
+
+if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
+  echo "Data Download"
+  for part in dev_set test_set train_set; do
+    local/download_and_untar.sh ${data_dir} ${data_url} ${part}
+  done
+fi
+
+if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
+  echo "Data preparation, prepare wav.scp/text/utt2spk/spk2utt"
+  for x in dev test train; do
+    mkdir -p data/$x
+    python local/prepare_data.py --src_dir $data_dir/$x --des_dir data/$x
+  done
+fi
+
+if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
+  echo "Extract campplus speaker embedding, you will get spk2embedding.pt and utt2embedding.pt in data/$x dir"
+  for x in dev test train; do
+    tools/extract_embedding.py --dir data/$x \
+      --onnx_path $pretrained_model_dir/campplus.onnx
+  done
+fi
+
+if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
+  echo "Extract discrete speech token, you will get utt2speech_token.pt in data/$x dir"
+  for x in dev test train; do
+    tools/extract_speech_token.py --dir data/$x \
+      --onnx_path $pretrained_model_dir/speech_tokenizer_v1.onnx
+  done
+fi
+
+if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
+  echo "Prepare required parquet format data, you should have prepared wav.scp/text/utt2spk/spk2utt/utt2embedding.pt/spk2embedding.pt/utt2speech_token.pt"
+  for x in dev test train; do
+    mkdir -p data/$x/parquet
+    tools/make_parquet_list.py --num_utts_per_parquet 1000 \
+      --num_processes 10 \
+      --src_dir data/$x \
+      --des_dir data/$x/parquet
+  done
+fi
+
+# inference
+if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
+  echo "Run inference. Please make sure utt in tts_text is in prompt_data"
+  for mode in sft zero_shot; do
+    python cosyvoice/bin/inference.py --mode $mode \
+      --gpu 0 \
+      --config conf/cosyvoice.yaml \
+      --prompt_data data/test/parquet/data.list \
+      --prompt_utt2data data/test/parquet/utt2data.list \
+      --tts_text `pwd`/tts_text.json \
+      --llm_model $pretrained_model_dir/llm.pt \
+      --flow_model $pretrained_model_dir/flow.pt \
+      --hifigan_model $pretrained_model_dir/hift.pt \
+      --result_dir `pwd`/exp/cosyvoice/test/$mode
+  done
+fi
+
+# train llm
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+num_gpus=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
+job_id=1986
+dist_backend="nccl"
+num_workers=2
+prefetch=100
+train_engine=torch_ddp
+if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
+  echo "Run train. We only support llm traning for now. If your want to train from scratch, please use conf/cosyvoice.fromscratch.yaml"
+  if [ $train_engine == 'deepspeed' ]; then
+    echo "Notice deepspeed has its own optimizer config. Modify conf/ds_stage2.json if necessary"
+  fi
+  cp data/train/parquet/data.list data/train.data.list
+  cp data/dev/parquet/data.list data/dev.data.list
+  for model in llm; do
+    torchrun --nnodes=1 --nproc_per_node=$num_gpus \
+        --rdzv_id=$job_id --rdzv_backend="c10d" --rdzv_endpoint="localhost:0" \
+      cosyvoice/bin/train.py \
+      --train_engine $train_engine \
+      --config conf/cosyvoice.yaml \
+      --train_data data/train.data.list \
+      --cv_data data/dev.data.list \
+      --model $model \
+      --checkpoint $pretrained_model_dir/$model.pt \
+      --model_dir `pwd`/exp/cosyvoice/$model/$train_engine \
+      --tensorboard_dir `pwd`/tensorboard/cosyvoice/$model/$train_engine \
+      --ddp.dist_backend $dist_backend \
+      --num_workers ${num_workers} \
+      --prefetch ${prefetch} \
+      --pin_memory \
+      --deepspeed_config ./conf/ds_stage2.json \
+      --deepspeed.save_states model+optimizer
+  done
+fi
\ No newline at end of file

examples/magicdata-read/cosyvoice/tools

+../../../tools
\ No newline at end of file

examples/magicdata-read/cosyvoice/tts_text.json

+{
+  "38_5718_20170915093303": [
+    "我想这出最好歌曲把歌词发到网上请别人帮我作曲急急",
+    "叫他明天早上差五分儿九点去机场"
+  ],
+  "38_5721_20170915091235": [
+    "变温室调到零下两度档",
+    "交谈中请勿轻信汇款信息陌生电话请勿使用外挂软件"
+  ],
+  "38_5733_20170915130323": [
+    "这是老鹰乐队的一首经典歌曲",
+    "我急用这段音乐我自己找到一段但是有现场杂音"
+  ],
+  "38_5836_20170916221414": [
+    "给我播一个陶喆的专辑",
+    "这套餐好贵呀我发这么多短信贵死了"
+  ]
+}
\ No newline at end of file

model.properties

+# 模型唯一标识
+modelCode = 919
+# 模型名称
+modelName=cosyvoice_pytorch                    
+# 模型描述
+modelDescription=CosyVoice多语言、音色和情感控制模型
+# 应用场景
+appScenario=推理,语音生成,金融,教育,交通,政府
+# 框架类型
+frameType=pytorch

requirements.txt

+conformer
+deepspeed
+diffusers
+gdown
+grpcio
+grpcio-tools
+hydra-core
+HyperPyYAML
+inflect
+librosa
+matplotlib
+modelscope
+networkx
+omegaconf
+pytorch-lightning
+onnxruntime
+openai-whisper
+protobuf
+pydantic
+rich
+soundfile
+tensorboard
+torch
+torchaudio
+wget
+fastapi
+fastapi-cli
+WeTextProcessing
+peft
+matcha-tts 
+gradio>=3.42.0
+pynini

runtime/python/Dockerfile

+FROM pytorch/pytorch:2.0.1-cuda11.7-cudnn8-runtime
+ENV DEBIAN_FRONTEND=noninteractive
+
+WORKDIR /opt/CosyVoice
+
+RUN sed -i s@/archive.ubuntu.com/@/mirrors.aliyun.com/@g /etc/apt/sources.list
+RUN apt-get update -y
+RUN apt-get -y install git unzip git-lfs
+RUN git lfs install
+RUN git clone --recursive https://github.com/FunAudioLLM/CosyVoice.git
+# here we use python==3.10 because we cannot find an image which have both python3.8 and torch2.0.1-cu118 installed
+RUN cd CosyVoice && pip3 install -r requirements.txt -i https://mirrors.aliyun.com/pypi/simple/ --trusted-host=mirrors.aliyun.com
+RUN cd CosyVoice/runtime/python/grpc && python3 -m grpc_tools.protoc -I. --python_out=. --grpc_python_out=. cosyvoice.proto
\ No newline at end of file

runtime/python/fastapi/client.py

+import argparse
+import logging
+import requests
+
+def saveResponse(path, response):
+    # 以二进制写入模式打开文件
+    with open(path, 'wb') as file:
+        # 将响应的二进制内容写入文件
+        file.write(response.content)
+
+def main():
+    api = args.api_base
+    if args.mode == 'sft':
+        url = api + "/api/inference/sft"
+        payload={
+            'tts': args.tts_text,
+            'role': args.spk_id
+        }
+        response = requests.request("POST", url, data=payload)
+        saveResponse(args.tts_wav, response)
+    elif args.mode == 'zero_shot':
+        url = api + "/api/inference/zero-shot"
+        payload={
+            'tts': args.tts_text,
+            'prompt': args.prompt_text
+        }
+        files=[('audio', ('prompt_audio.wav', open(args.prompt_wav,'rb'), 'application/octet-stream'))]
+        response = requests.request("POST", url, data=payload, files=files)
+        saveResponse(args.tts_wav, response)
+    elif args.mode == 'cross_lingual':
+        url = api + "/api/inference/cross-lingual"
+        payload={
+            'tts': args.tts_text,
+        }
+        files=[('audio', ('prompt_audio.wav', open(args.prompt_wav,'rb'), 'application/octet-stream'))]
+        response = requests.request("POST", url, data=payload, files=files)
+        saveResponse(args.tts_wav, response)
+    else:
+        url = api + "/api/inference/instruct"
+        payload = {
+            'tts': args.tts_text,
+            'role': args.spk_id,
+            'instruct': args.instruct_text
+        }
+        response = requests.request("POST", url, data=payload)
+        saveResponse(args.tts_wav, response)
+    logging.info("Response save to {}", args.tts_wav)
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--api_base',
+                        type=str,
+                        default='http://127.0.0.1:6006')
+    parser.add_argument('--mode',
+                        default='sft',
+                        choices=['sft', 'zero_shot', 'cross_lingual', 'instruct'],
+                        help='request mode')
+    parser.add_argument('--tts_text',
+                        type=str,
+                        default='你好，我是通义千问语音合成大模型，请问有什么可以帮您的吗？')
+    parser.add_argument('--spk_id',
+                        type=str,
+                        default='中文女')
+    parser.add_argument('--prompt_text',
+                        type=str,
+                        default='希望你以后能够做的比我还好呦。')
+    parser.add_argument('--prompt_wav',
+                        type=str,
+                        default='../../../zero_shot_prompt.wav')
+    parser.add_argument('--instruct_text',
+                        type=str,
+                        default='Theo \'Crimson\', is a fiery, passionate rebel leader. Fights with fervor for justice, but struggles with impulsiveness.')
+    parser.add_argument('--tts_wav',
+                        type=str,
+                        default='demo.wav')
+    args = parser.parse_args()
+    prompt_sr, target_sr = 16000, 22050
+    main()

runtime/python/fastapi/server.py

+# Set inference model
+# export MODEL_DIR=pretrained_models/CosyVoice-300M-Instruct
+# For development
+# fastapi dev --port 6006 fastapi_server.py
+# For production deployment
+# fastapi run --port 6006 fastapi_server.py
+
+import os
+import sys
+import io,time
+from fastapi import FastAPI, Response, File, UploadFile, Form
+from fastapi.responses import HTMLResponse
+from fastapi.middleware.cors import CORSMiddleware  #引入 CORS中间件模块
+from contextlib import asynccontextmanager
+ROOT_DIR = os.path.dirname(os.path.abspath(__file__))
+sys.path.append('{}/../../..'.format(ROOT_DIR))
+sys.path.append('{}/../../../third_party/Matcha-TTS'.format(ROOT_DIR))
+from cosyvoice.cli.cosyvoice import CosyVoice
+from cosyvoice.utils.file_utils import load_wav
+import numpy as np
+import torch
+import torchaudio
+import logging
+logging.getLogger('matplotlib').setLevel(logging.WARNING)
+
+class LaunchFailed(Exception):
+    pass
+
+@asynccontextmanager
+async def lifespan(app: FastAPI):
+    model_dir = os.getenv("MODEL_DIR", "pretrained_models/CosyVoice-300M-SFT")
+    if model_dir:
+        logging.info("MODEL_DIR is {}", model_dir)
+        app.cosyvoice = CosyVoice(model_dir)
+        # sft usage
+        logging.info("Avaliable speakers {}", app.cosyvoice.list_avaliable_spks())
+    else:
+        raise LaunchFailed("MODEL_DIR environment must set")
+    yield
+
+app = FastAPI(lifespan=lifespan)
+
+#设置允许访问的域名
+origins = ["*"]  #"*"，即为所有,也可以改为允许的特定ip。
+app.add_middleware(
+    CORSMiddleware, 
+    allow_origins=origins,  #设置允许的origins来源
+    allow_credentials=True,
+    allow_methods=["*"],  # 设置允许跨域的http方法，比如 get、post、put等。
+    allow_headers=["*"])  #允许跨域的headers，可以用来鉴别来源等作用。
+
+def buildResponse(output):
+    buffer = io.BytesIO()
+    torchaudio.save(buffer, output, 22050, format="wav")
+    buffer.seek(0)
+    return Response(content=buffer.read(-1), media_type="audio/wav")
+
+@app.post("/api/inference/sft")
+@app.get("/api/inference/sft")
+async def sft(tts: str = Form(), role: str = Form()):
+    start = time.process_time()
+    output = app.cosyvoice.inference_sft(tts, role)
+    end = time.process_time()
+    logging.info("infer time is {} seconds", end-start)
+    return buildResponse(output['tts_speech'])
+
+@app.post("/api/inference/zero-shot")
+async def zeroShot(tts: str = Form(), prompt: str = Form(), audio: UploadFile = File()):
+    start = time.process_time()
+    prompt_speech = load_wav(audio.file, 16000)
+    prompt_audio = (prompt_speech.numpy() * (2**15)).astype(np.int16).tobytes()
+    prompt_speech_16k = torch.from_numpy(np.array(np.frombuffer(prompt_audio, dtype=np.int16))).unsqueeze(dim=0)
+    prompt_speech_16k = prompt_speech_16k.float() / (2**15)
+
+    output = app.cosyvoice.inference_zero_shot(tts, prompt, prompt_speech_16k)
+    end = time.process_time()
+    logging.info("infer time is {} seconds", end-start)
+    return buildResponse(output['tts_speech'])
+
+@app.post("/api/inference/cross-lingual")
+async def crossLingual(tts: str = Form(), audio: UploadFile = File()):
+    start = time.process_time()
+    prompt_speech = load_wav(audio.file, 16000)
+    prompt_audio = (prompt_speech.numpy() * (2**15)).astype(np.int16).tobytes()
+    prompt_speech_16k = torch.from_numpy(np.array(np.frombuffer(prompt_audio, dtype=np.int16))).unsqueeze(dim=0)
+    prompt_speech_16k = prompt_speech_16k.float() / (2**15)
+
+    output = app.cosyvoice.inference_cross_lingual(tts, prompt_speech_16k)
+    end = time.process_time()
+    logging.info("infer time is {} seconds", end-start)
+    return buildResponse(output['tts_speech'])
+
+@app.post("/api/inference/instruct")
+@app.get("/api/inference/instruct")
+async def instruct(tts: str = Form(), role: str = Form(), instruct: str = Form()):
+    start = time.process_time()
+    output = app.cosyvoice.inference_instruct(tts, role, instruct)
+    end = time.process_time()
+    logging.info("infer time is {} seconds", end-start)
+    return buildResponse(output['tts_speech'])
+
+@app.get("/api/roles")
+async def roles():
+    return {"roles": app.cosyvoice.list_avaliable_spks()}
+
+@app.get("/", response_class=HTMLResponse)
+async def root():
+    return """
+    <!DOCTYPE html>
+    <html lang=zh-cn>
+        <head>
+            <meta charset=utf-8>
+            <title>Api information</title>
+        </head>
+        <body>
+            Get the supported tones from the Roles API first, then enter the tones and textual content in the TTS API for synthesis. <a href='./docs'>Documents of API</a>
+        </body>
+    </html>
+    """

runtime/python/grpc/client.py

+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import os
+import sys
+ROOT_DIR = os.path.dirname(os.path.abspath(__file__))
+sys.path.append('{}/../../..'.format(ROOT_DIR))
+sys.path.append('{}/../../../third_party/Matcha-TTS'.format(ROOT_DIR))
+import logging
+import argparse
+import torchaudio
+import cosyvoice_pb2
+import cosyvoice_pb2_grpc
+import grpc
+import torch
+import numpy as np
+from cosyvoice.utils.file_utils import load_wav
+
+
+def main():
+    with grpc.insecure_channel("{}:{}".format(args.host, args.port)) as channel:
+        stub = cosyvoice_pb2_grpc.CosyVoiceStub(channel)
+        request = cosyvoice_pb2.Request()
+        if args.mode == 'sft':
+            logging.info('send sft request')
+            sft_request = cosyvoice_pb2.sftRequest()
+            sft_request.spk_id = args.spk_id
+            sft_request.tts_text = args.tts_text
+            request.sft_request.CopyFrom(sft_request)
+        elif args.mode == 'zero_shot':
+            logging.info('send zero_shot request')
+            zero_shot_request = cosyvoice_pb2.zeroshotRequest()
+            zero_shot_request.tts_text = args.tts_text
+            zero_shot_request.prompt_text = args.prompt_text
+            prompt_speech = load_wav(args.prompt_wav, 16000)
+            zero_shot_request.prompt_audio = (prompt_speech.numpy() * (2**15)).astype(np.int16).tobytes()
+            request.zero_shot_request.CopyFrom(zero_shot_request)
+        elif args.mode == 'cross_lingual':
+            logging.info('send cross_lingual request')
+            cross_lingual_request = cosyvoice_pb2.crosslingualRequest()
+            cross_lingual_request.tts_text = args.tts_text
+            prompt_speech = load_wav(args.prompt_wav, 16000)
+            cross_lingual_request.prompt_audio = (prompt_speech.numpy() * (2**15)).astype(np.int16).tobytes()
+            request.cross_lingual_request.CopyFrom(cross_lingual_request)
+        else:
+            logging.info('send instruct request')
+            instruct_request = cosyvoice_pb2.instructRequest()
+            instruct_request.tts_text = args.tts_text
+            instruct_request.spk_id = args.spk_id
+            instruct_request.instruct_text = args.instruct_text
+            request.instruct_request.CopyFrom(instruct_request)
+
+        response = stub.Inference(request)
+        logging.info('save response to {}'.format(args.tts_wav))
+        tts_speech = torch.from_numpy(np.array(np.frombuffer(response.tts_audio, dtype=np.int16))).unsqueeze(dim=0)
+        torchaudio.save(args.tts_wav, tts_speech, target_sr)
+        logging.info('get response')
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--host',
+                        type=str,
+                        default='0.0.0.0')
+    parser.add_argument('--port',
+                        type=int,
+                        default='50000')
+    parser.add_argument('--mode',
+                        default='sft',
+                        choices=['sft', 'zero_shot', 'cross_lingual', 'instruct'],
+                        help='request mode')
+    parser.add_argument('--tts_text',
+                        type=str,
+                        default='你好，我是通义千问语音合成大模型，请问有什么可以帮您的吗？')
+    parser.add_argument('--spk_id',
+                        type=str,
+                        default='中文女')
+    parser.add_argument('--prompt_text',
+                        type=str,
+                        default='希望你以后能够做的比我还好呦。')
+    parser.add_argument('--prompt_wav',
+                        type=str,
+                        default='../../../zero_shot_prompt.wav')
+    parser.add_argument('--instruct_text',
+                        type=str,
+                        default='Theo \'Crimson\', is a fiery, passionate rebel leader. Fights with fervor for justice, but struggles with impulsiveness.')
+    parser.add_argument('--tts_wav',
+                        type=str,
+                        default='demo.wav')
+    args = parser.parse_args()
+    prompt_sr, target_sr = 16000, 22050
+    main()

runtime/python/grpc/cosyvoice.proto

+syntax = "proto3";
+
+package cosyvoice;
+option go_package = "protos/";
+
+service CosyVoice{
+  rpc Inference(Request) returns (Response) {}
+}
+
+message Request{
+  oneof RequestPayload {
+    sftRequest sft_request = 1;
+    zeroshotRequest zero_shot_request = 2;
+    crosslingualRequest cross_lingual_request = 3;
+    instructRequest instruct_request = 4;
+  }
+}
+
+message sftRequest{
+  string spk_id = 1;
+  string tts_text = 2;
+}
+
+message zeroshotRequest{
+  string tts_text = 1;
+  string prompt_text = 2;
+  bytes prompt_audio = 3;
+}
+
+message crosslingualRequest{
+  string tts_text = 1;
+  bytes prompt_audio = 2;
+}
+
+message instructRequest{
+  string tts_text = 1;
+  string spk_id = 2;
+  string instruct_text = 3;
+}
+
+message Response{
+  bytes tts_audio = 1;
+}
\ No newline at end of file

runtime/python/grpc/server.py

+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import os
+import sys
+ROOT_DIR = os.path.dirname(os.path.abspath(__file__))
+sys.path.append('{}/../../..'.format(ROOT_DIR))
+sys.path.append('{}/../../../third_party/Matcha-TTS'.format(ROOT_DIR))
+from concurrent import futures
+import argparse
+import cosyvoice_pb2
+import cosyvoice_pb2_grpc
+import logging
+logging.getLogger('matplotlib').setLevel(logging.WARNING)
+import grpc
+import torch
+import numpy as np
+from cosyvoice.cli.cosyvoice import CosyVoice
+
+logging.basicConfig(level=logging.DEBUG,
+                    format='%(asctime)s %(levelname)s %(message)s')
+
+class CosyVoiceServiceImpl(cosyvoice_pb2_grpc.CosyVoiceServicer):
+    def __init__(self, args):
+        self.cosyvoice = CosyVoice(args.model_dir)
+        logging.info('grpc service initialized')
+
+    def Inference(self, request, context):
+        if request.HasField('sft_request'):
+            logging.info('get sft inference request')
+            model_output = self.cosyvoice.inference_sft(request.sft_request.tts_text, request.sft_request.spk_id)
+        elif request.HasField('zero_shot_request'):
+            logging.info('get zero_shot inference request')
+            prompt_speech_16k = torch.from_numpy(np.array(np.frombuffer(request.zero_shot_request.prompt_audio, dtype=np.int16))).unsqueeze(dim=0)
+            prompt_speech_16k = prompt_speech_16k.float() / (2**15)
+            model_output = self.cosyvoice.inference_zero_shot(request.zero_shot_request.tts_text, request.zero_shot_request.prompt_text, prompt_speech_16k)
+        elif request.HasField('cross_lingual_request'):
+            logging.info('get cross_lingual inference request')
+            prompt_speech_16k = torch.from_numpy(np.array(np.frombuffer(request.cross_lingual_request.prompt_audio, dtype=np.int16))).unsqueeze(dim=0)
+            prompt_speech_16k = prompt_speech_16k.float() / (2**15)
+            model_output = self.cosyvoice.inference_cross_lingual(request.cross_lingual_request.tts_text, prompt_speech_16k)
+        else:
+            logging.info('get instruct inference request')
+            model_output = self.cosyvoice.inference_instruct(request.instruct_request.tts_text, request.instruct_request.spk_id, request.instruct_request.instruct_text)
+
+        logging.info('send inference response')
+        response = cosyvoice_pb2.Response()
+        response.tts_audio = (model_output['tts_speech'].numpy() * (2 ** 15)).astype(np.int16).tobytes()
+        return response
+
+def main():
+    grpcServer = grpc.server(futures.ThreadPoolExecutor(max_workers=args.max_conc), maximum_concurrent_rpcs=args.max_conc)
+    cosyvoice_pb2_grpc.add_CosyVoiceServicer_to_server(CosyVoiceServiceImpl(args), grpcServer)
+    grpcServer.add_insecure_port('0.0.0.0:{}'.format(args.port))
+    grpcServer.start()
+    logging.info("server listening on 0.0.0.0:{}".format(args.port))
+    grpcServer.wait_for_termination()
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--port',
+                        type=int,
+                        default=50000)
+    parser.add_argument('--max_conc',
+                        type=int,
+                        default=4)
+    parser.add_argument('--model_dir',
+                        type=str,
+                        default='iic/CosyVoice-300M',
+                        help='local path or modelscope repo id')
+    args = parser.parse_args()
+    main()

tools/extract_embedding.py

+#!/usr/bin/env python3
+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import argparse
+import torch
+import torchaudio
+from tqdm import tqdm
+import onnxruntime
+import torchaudio.compliance.kaldi as kaldi
+
+
+def main(args):
+    utt2wav, utt2spk = {}, {}
+    with open('{}/wav.scp'.format(args.dir)) as f:
+        for l in f:
+            l = l.replace('\n', '').split()
+            utt2wav[l[0]] = l[1]
+    with open('{}/utt2spk'.format(args.dir)) as f:
+        for l in f:
+            l = l.replace('\n', '').split()
+            utt2spk[l[0]] = l[1]
+
+    option = onnxruntime.SessionOptions()
+    option.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
+    option.intra_op_num_threads = 1
+    providers = ["CPUExecutionProvider"]
+    ort_session = onnxruntime.InferenceSession(args.onnx_path, sess_options=option, providers=providers)
+
+    utt2embedding, spk2embedding = {}, {}
+    for utt in tqdm(utt2wav.keys()):
+        audio, sample_rate = torchaudio.load(utt2wav[utt])
+        if sample_rate != 16000:
+            audio = torchaudio.transforms.Resample(orig_freq=sample_rate, new_freq=16000)(audio)
+        feat = kaldi.fbank(audio,
+                           num_mel_bins=80,
+                           dither=0,
+                           sample_frequency=16000)
+        feat = feat - feat.mean(dim=0, keepdim=True)
+        embedding = ort_session.run(None, {ort_session.get_inputs()[0].name: feat.unsqueeze(dim=0).cpu().numpy()})[0].flatten().tolist()
+        utt2embedding[utt] = embedding
+        spk = utt2spk[utt]
+        if spk not in spk2embedding:
+            spk2embedding[spk] = []
+        spk2embedding[spk].append(embedding)
+    for k, v in spk2embedding.items():
+        spk2embedding[k] = torch.tensor(v).mean(dim=0).tolist()
+
+    torch.save(utt2embedding, '{}/utt2embedding.pt'.format(args.dir))
+    torch.save(spk2embedding, '{}/spk2embedding.pt'.format(args.dir))
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--dir',
+                        type=str)
+    parser.add_argument('--onnx_path',
+                        type=str)
+    args = parser.parse_args()
+    main(args)