icefall/egs/speech_llm/SPEECH2SPEECH/prepare.sh

#!/usr/bin/env bash

# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python

export PYTHONPATH=$PYTHONPATH:/workspace/icefall

set -eou pipefail

stage=$1
stop_stage=$2
# All files generated by this script are saved in "data".
# You can safely remove "data" and rerun this script to regenerate it.
mkdir -p data

log() {
  # This function is from espnet
  local fname=${BASH_SOURCE[1]##*/}
  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
}


if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
  log "stage 0: Clone CosyVoice repo and install requirements inside the container"
  # docker: ghcr.io/swivid/f5-tts:main
  pip install k2==1.24.4.dev20241030+cuda12.4.torch2.4.0 -f https://k2-fsa.github.io/k2/cuda.html
  git clone --recursive https://github.com/FunAudioLLM/CosyVoice.git /workspace/CosyVoice
  cd /workspace/CosyVoice
  # If you failed to clone submodule due to network failures, please run following command until success
  git submodule update --init --recursive
  pip install -r qwen_omni/requirements.txt
  pip install -r qwen_omni/requirements-cosyvoice.txt

  # For Chinese only dataset, you can use the following command to download the Chinese fine-tuned whisper model.
  huggingface-cli download --local-dir models/whisper yuekai/icefall_asr_multi-hans-zh_whisper
  # Cosyvoice pretrained model for speech token2wav module
  huggingface-cli download --local-dir models/CosyVoice-300M-SFT FunAudioLLM/CosyVoice-300M-SFT
  # Qwen Pretrained model
  huggingface-cli download --local-dir models/Qwen2.5-0.5B-Instruct Qwen/Qwen2.5-0.5B-Instruct
  # Qwen-Omni like speech2speech model trained on worstchan/Belle_1.4M-SLAM-Omni
  huggingface-cli download --local-dir models/qwen-omni-like-speech2speech-belle-1.4M yuekai/qwen-omni-like-speech2speech-belle-1.4M

  # For Gradio demo, we follow https://arxiv.org/abs/2412.15649 to use ASR model to decode the history speech as context.
  pip install sherpa-onnx
  model_path=local/sherpa-onnx-paraformer-zh-2023-09-14
  if [ ! -d $model_path ]; then
    wget -nc https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-paraformer-zh-2023-09-14.tar.bz2
    tar xvf sherpa-onnx-paraformer-zh-2023-09-14.tar.bz2 -C local
  fi
fi
export PYTHONPATH=$PYTHONPATH:/workspace/CosyVoice

if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
  log "stage 1: Compute fbank feature from huggingface"
  python3 local/compute_whisper_fbank.py \
   --num-mel-bins 80 --whisper-fbank True --resample-to-16kHz True --speed-perturb False \
   --out-dir data/fbank_test \
   --huggingface-dataset-path-or-name /workspace/Belle_1.4M-SLAM-Omni \
   --audio-key question_audio --text-key answer \
   --prefix belle
fi

if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
  log "Stage 2: Combine features"
  manifest_dir=data/fbank
  if [ ! -f $manifest_dir/cuts_belle_00001-01600.jsonl.gz ]; then
    mv $manifest_dir/cuts_belle.00000.jsonl.gz ./
    # exclude cust_belle_00000.jsonl.gz for valid and test set
    pieces=$(find $manifest_dir -name "cuts_belle.*.jsonl.gz" | sort)
    echo $pieces | wc
    lhotse combine $pieces data/fbank/cuts_belle_00001-01600.jsonl.gz
    mv ./cuts_belle.00000.jsonl.gz $manifest_dir # put it back
    cd $manifest_dir && ln -s cuts_belle_00001-01600.jsonl.gz cuts_belle_train.jsonl.gz
    ln -s cuts_belle.00000.jsonl.gz cuts_belle_test.jsonl.gz && cd -
  fi
fi

ngpu=8
exp_dir=./qwen_omni/exp_speech2speech
if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
  log "stage 3: Training Speech2Speech Model"
  torchrun --nproc_per_node $ngpu ./qwen_omni/train.py \
    --max-duration 50 \
    --enable-musan False \
    --exp-dir $exp_dir \
    --speech-encoder-path-or-name models/whisper/v1.1/whisper-large-v2-multi-hans-zh-epoch-3-avg-10.pt \
    --llm-path-or-name Qwen/Qwen2.5-0.5B-Instruct \
    --manifest-dir data/fbank \
    --deepspeed \
    --deepspeed_config ./qwen_omni/ds_config_zero1.json \
    --use-flash-attn True \
    --use-lora True --unfreeze-llm True --unfreeze-speech-projector True --enable-speech-output True
fi

if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
  log "stage 4: Decoding, only support batch_size=1 for now."
  cd $exp_dir && ln -s ../../models/qwen-omni-like-speech2speech-belle-1.4M/pytorch_model.bin epoch-999.pt && cd -
  python3 ./qwen_omni/decode.py \
    --max-duration 1 \
    --exp-dir $exp_dir \
    --speech-encoder-path-or-name models/whisper/v1.1/whisper-large-v2-multi-hans-zh-epoch-3-avg-10.pt  \
    --llm-path-or-name models/Qwen2.5-0.5B-Instruct \
    --epoch 999 --avg 1 \
    --manifest-dir data/fbank \
    --use-flash-attn True \
    --method e2e-epoch10_speech2speech \
    --enable-speech-output True \
    --token2wav-path models/CosyVoice-300M-SFT \
    --use-lora True
fi

if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
  log "stage 5: Gradio Demo"
  python3 ./qwen_omni/web_demo.py \
    --speech-encoder-path-or-name models/whisper/v1.1/whisper-large-v2-multi-hans-zh-epoch-3-avg-10.pt  \
    --llm-path-or-name models/Qwen2.5-0.5B-Instruct \
    --checkpoint-path $exp_dir/epoch-999.pt \
    --use-flash-attn True \
    --enable-speech-output True \
    --asr-model-dir local/sherpa-onnx-paraformer-zh-2023-09-14 \
    --use-lora True --token2wav-path /workspace/CosyVoice-300M-SFT --share
fi

if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
  log "stage 1: Compute fbank feature from huggingface"
  # CUDA_VISIBLE_DEVICES=0 python3 local/compute_whisper_fbank.py \
  #  --num-mel-bins 80 --whisper-fbank True --resample-to-16kHz True --speed-perturb False \
  #  --out-dir data/fbank_voice_assistant \
  #  --huggingface-dataset-path-or-name worstchan/VoiceAssistant-400K-SLAM-Omni \
  #  --audio-key question_audio --text-key answer \
  #  --prefix voice_assistant
  CUDA_VISIBLE_DEVICES=0 python3 local/compute_whisper_fbank.py \
   --num-mel-bins 80 --whisper-fbank True --resample-to-16kHz True --speed-perturb False \
   --out-dir data/fbank_voice_assistant_cosy2 \
   --json-file-path /workspace/slam/VoiceAssistant-430K-vocalnet/VoiceAssistant-430K.json \
   --prefix voice_assistant
fi

if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
  log "stage 7: Compute fbank feature from huggingface"
  # CUDA_VISIBLE_DEVICES=1 python3 local/compute_whisper_fbank.py \
  #  --num-mel-bins 80 --whisper-fbank True --resample-to-16kHz True --speed-perturb False \
  #  --out-dir data/fbank_ultrachat \
  #  --huggingface-dataset-path-or-name worstchan/UltraChat-300K-SLAM-Omni \
  #  --audio-key question_audio --text-key answer \
  #  --prefix ultrachat
  CUDA_VISIBLE_DEVICES=1 python3 local/compute_whisper_fbank.py \
   --num-mel-bins 80 --whisper-fbank True --resample-to-16kHz True --speed-perturb False \
   --out-dir data/fbank_ultrachat_cosy2 \
   --json-file-path /workspace/slam/UltraChat-vocalnet/UltraChat.json \
   --prefix ultrachat
fi

if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then
  log "stage 8: Compute fbank feature from huggingface"

  CUDA_VISIBLE_DEVICES=1 python3 local/compute_whisper_fbank.py \
   --num-mel-bins 80 --whisper-fbank True --resample-to-16kHz True --speed-perturb False \
   --out-dir data/fbank_gigaspeech \
   --huggingface-dataset-path-or-name speechcolab/gigaspeech \
   --subset test --split test \
   --audio-key audio --text-key text \
   --prefix gigaspeech
fi

if [ $stage -le 9 ] && [ $stop_stage -ge 9 ]; then
  log "stage 9: Compute fbank feature from huggingface"
  CUDA_VISIBLE_DEVICES=0 python3 local/compute_whisper_fbank.py \
   --num-mel-bins 80 --whisper-fbank True --resample-to-16kHz True --speed-perturb True \
   --out-dir data/fbank_gigaspeech \
   --huggingface-dataset-path-or-name speechcolab/gigaspeech \
   --subset xl --split train \
   --audio-key audio --text-key text \
   --prefix gigaspeech
fi

# cd /workspace && ln -s /lustre/fsw/general_sa/yuekaiz/s2s slam && cd -
ngpu=4
exp_dir=./qwen_omni/exp_speech2speech_en
if [ $stage -le 10 ] && [ $stop_stage -ge 10 ]; then
  log "stage 10: Training Speech2Speech Model"
  torchrun --nproc_per_node $ngpu ./qwen_omni/train.py \
    --max-duration 150 \
    --enable-musan False \
    --exp-dir $exp_dir \
    --speech-encoder-path-or-name models/large-v2.pt \
    --llm-path-or-name Qwen/Qwen2.5-0.5B-Instruct \
    --dataset-format vocalnet \
    --manifest-dir data/fbank \
    --deepspeed \
    --deepspeed_config ./qwen_omni/ds_config_zero1.json \
    --use-flash-attn True --on-the-fly-feats True \
    --use-lora True --unfreeze-llm True --unfreeze-speech-projector True --enable-speech-output True
fi


if [ $stage -le 11 ] && [ $stop_stage -ge 11 ]; then
  log "stage 11: Decoding EN, only support batch_size=1 for now."
  exp_dir=./qwen_omni/exp_speech2speech_en_continue
  # cd $exp_dir && ln -s ../../models/qwen-omni-like-speech2speech-belle-1.4M/pytorch_model.bin epoch-999.pt && cd -
  python3 ./qwen_omni/decode.py \
    --max-duration 1 \
    --exp-dir $exp_dir \
    --speech-encoder-path-or-name models/large-v2.pt  \
    --llm-path-or-name models/Qwen2.5-0.5B-Instruct \
    --epoch 997 --avg 1 \
    --manifest-dir data/fbank \
    --use-flash-attn True \
    --method e2e-epoch4_speech2speech \
    --enable-speech-output True \
    --token2wav-path /workspace/CosyVoice2-0.5B \
    --use-lora True
fi