init

2025-06-08 16:07:40 +08:00
commit 6840820f8d
11 changed files with 1023 additions and 0 deletions
@@ -0,0 +1,26 @@
+name: Publish to Comfy registry
+on:
+  workflow_dispatch:
+  push:
+    branches:
+      - master
+      - main
+    paths:
+      - "pyproject.toml"
+
+permissions:
+  issues: write
+
+jobs:
+  publish-node:
+    name: Publish Custom Node to registry
+    runs-on: ubuntu-latest
+    if: ${{ github.repository_owner == 'avenstack' }}
+    steps:
+      - name: Check out code
+        uses: actions/checkout@v4
+      - name: Publish Custom Node
+        uses: Comfy-Org/publish-node-action@v1
+        with:
+          ## Add your own personal access token to your Github Repository secrets and reference it here.
+          personal_access_token: ${{ secrets.REGISTRY_ACCESS_TOKEN }}
@@ -0,0 +1,3 @@
+*.pyc
+__pycache__/
+.hisotry/
@@ -0,0 +1,63 @@
+## ComfyUI FunAsr Nodes
+
+使用FunAsr进行语音识别，并转为文本或字幕文件。
+
+![workflow](example_workflow/workflow.webp)
+
+## 功能
+
+- [x] 语音识别
+- [x] 语音时间戳预测
+- [x] 语音转字幕
+
+## 安装
+
+```
+cd ComfyUI/custom_nodes
+git clone https://github.com/avenstack/ComfyUI-AV-FunASR.git
+cd ComfyUI-AV-FunASR
+pip install -r requirements.txt
+```
+
+## 模型下载
+
+1. [语音识别](https://modelscope.cn/models/iic/speech_seaco_paraformer_large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/files)
+2. [语音端点检测](modelscope.cn/models/iic/speech_fsmn_vad_zh-cn-16k-common-pytorch/files)
+3. [语音时间戳预测](modelscope.cn/models/iic/speech_timestamp_prediction-v1-16k-offline)
+
+模型存放目录：`models/ASR/FunASR/iic`
+```
+ iic
+    ├── speech_fsmn_vad_zh-cn-16k-common-pytorch
+    │   ├── README.md
+    │   ├── am.mvn
+    │   ├── config.yaml
+    │   ├── configuration.json
+    │   ├── example
+    │   ├── fig
+    │   └── model.pt
+    ├── speech_seaco_paraformer_large_asr_nat-zh-cn-16k-common-vocab8404-pytorch
+    │   ├── README.md
+    │   ├── am.mvn
+    │   ├── asr_example_hotword.wav
+    │   ├── config.yaml
+    │   ├── configuration.json
+    │   ├── example
+    │   ├── fig
+    │   ├── model.pt
+    │   ├── seg_dict
+    │   └── tokens.json
+    └── speech_timestamp_prediction-v1-16k-offline
+        ├── README.md
+        ├── am.mvn
+        ├── config.yaml
+        ├── configuration.json
+        ├── example
+        ├── model.pt
+        ├── seg_dict
+        └── tokens.json
+```
+
+## 鸣谢
+
+- [ComfyUI](https://github.com/comfyanonymous/ComfyUI)
@@ -0,0 +1,3 @@
+from .nodes import NODE_CLASS_MAPPINGS, NODE_DISPLAY_NAME_MAPPINGS
+
+__all__ = ['NODE_CLASS_MAPPINGS', 'NODE_DISPLAY_NAME_MAPPINGS']
@@ -0,0 +1,311 @@
+{
+  "id": "377e0192-2ad2-408a-871c-6a86cf058f73",
+  "revision": 0,
+  "last_node_id": 33,
+  "last_link_id": 50,
+  "nodes": [
+    {
+      "id": 2,
+      "type": "LoadAudio",
+      "pos": [
+        182.18014526367188,
+        389.2699890136719
+      ],
+      "size": [
+        270,
+        136
+      ],
+      "flags": {},
+      "order": 0,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "AUDIO",
+          "type": "AUDIO",
+          "links": [
+            24
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.3.32",
+        "Node name for S&R": "LoadAudio"
+      },
+      "widgets_values": [
+        "4_jidaohuman-megatts3_0.flac",
+        null,
+        null
+      ]
+    },
+    {
+      "id": 25,
+      "type": "PrimitiveStringMultiline",
+      "pos": [
+        186.9199676513672,
+        591.9599609375
+      ],
+      "size": [
+        260,
+        163
+      ],
+      "flags": {
+        "collapsed": false
+      },
+      "order": 1,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "STRING",
+          "type": "STRING",
+          "links": [
+            47,
+            48
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.3.32",
+        "Node name for S&R": "PrimitiveStringMultiline"
+      },
+      "widgets_values": [
+        "直观来说，测试时试错的步骤只有当能带来信息增益的情况下才有益，然而人们并没有在RL训练中告诉模型试错和反思带来的信息增益。"
+      ]
+    },
+    {
+      "id": 18,
+      "type": "AVSpeechTimestamp",
+      "pos": [
+        489.56988525390625,
+        374.6600341796875
+      ],
+      "size": [
+        210,
+        128
+      ],
+      "flags": {
+        "collapsed": false
+      },
+      "order": 2,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "audio",
+          "type": "AUDIO",
+          "link": 24
+        },
+        {
+          "name": "text",
+          "type": "STRING",
+          "widget": {
+            "name": "text"
+          },
+          "link": 47
+        }
+      ],
+      "outputs": [
+        {
+          "name": "text",
+          "type": "STRING",
+          "links": []
+        },
+        {
+          "name": "json_result",
+          "type": "STRING",
+          "links": []
+        },
+        {
+          "name": "asr_result",
+          "type": "ASRRESULT",
+          "links": [
+            43
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "AVSpeechTimestamp"
+      },
+      "widgets_values": [
+        "",
+        false
+      ]
+    },
+    {
+      "id": 30,
+      "type": "AVFormat2Subtitle",
+      "pos": [
+        518.5700073242188,
+        561.3599853515625
+      ],
+      "size": [
+        210,
+        88
+      ],
+      "flags": {
+        "collapsed": false
+      },
+      "order": 3,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "asr_result",
+          "type": "ASRRESULT",
+          "link": 43
+        },
+        {
+          "name": "text",
+          "type": "STRING",
+          "widget": {
+            "name": "text"
+          },
+          "link": 48
+        }
+      ],
+      "outputs": [
+        {
+          "name": "subtitle",
+          "type": "STRING",
+          "links": [
+            45,
+            50
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "AVFormat2Subtitle"
+      },
+      "widgets_values": [
+        ""
+      ]
+    },
+    {
+      "id": 31,
+      "type": "AVSaveSubtitles",
+      "pos": [
+        778.79541015625,
+        357.02178955078125
+      ],
+      "size": [
+        270,
+        82
+      ],
+      "flags": {
+        "collapsed": true
+      },
+      "order": 4,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "subtitles",
+          "type": "STRING",
+          "widget": {
+            "name": "subtitles"
+          },
+          "link": 45
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "Node name for S&R": "AVSaveSubtitles"
+      },
+      "widgets_values": [
+        "",
+        "subtitles"
+      ]
+    },
+    {
+      "id": 33,
+      "type": "Bjornulf_ShowText",
+      "pos": [
+        814.5744018554688,
+        427.3533935546875
+      ],
+      "size": [
+        650.6199951171875,
+        437.7900085449219
+      ],
+      "flags": {},
+      "order": 5,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "text_value",
+          "type": "STRING",
+          "link": 50
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "cnr_id": "bjornulf_custom_nodes",
+        "ver": "1.0.0",
+        "Node name for S&R": "Bjornulf_ShowText"
+      },
+      "widgets_values": [
+        "1\n00:00:00,170 --> 00:00:02,330\n直观来说 测试时试错的步骤\n\n2\n00:00:02,330 --> 00:00:04,590\n只有当能带来信息增益的情况下才有益\n\n3\n00:00:04,590 --> 00:00:06,870\n然而人们并没有在RL训练\n\n4\n00:00:06,870 --> 00:00:09,330\n中告诉模型试错和反思带来的信息增益\n\n5\n00:00:09,330 --> 00:00:09,470\n\n\n"
+      ]
+    }
+  ],
+  "links": [
+    [
+      24,
+      2,
+      0,
+      18,
+      0,
+      "AUDIO"
+    ],
+    [
+      43,
+      18,
+      2,
+      30,
+      0,
+      "ASRRESULT"
+    ],
+    [
+      45,
+      30,
+      0,
+      31,
+      0,
+      "STRING"
+    ],
+    [
+      47,
+      25,
+      0,
+      18,
+      1,
+      "STRING"
+    ],
+    [
+      48,
+      25,
+      0,
+      30,
+      1,
+      "STRING"
+    ],
+    [
+      50,
+      30,
+      0,
+      33,
+      0,
+      "STRING"
+    ]
+  ],
+  "groups": [],
+  "config": {},
+  "extra": {
+    "frontendVersion": "1.18.9",
+    "VHS_latentpreview": false,
+    "VHS_latentpreviewrate": 0,
+    "VHS_MetadataImage": true,
+    "VHS_KeepIntermediate": true
+  },
+  "version": 0.4
+}
@@ -0,0 +1,241 @@
+{
+  "id": "377e0192-2ad2-408a-871c-6a86cf058f73",
+  "revision": 0,
+  "last_node_id": 35,
+  "last_link_id": 53,
+  "nodes": [
+    {
+      "id": 2,
+      "type": "LoadAudio",
+      "pos": [
+        187.18014526367188,
+        368.2699890136719
+      ],
+      "size": [
+        270,
+        136
+      ],
+      "flags": {},
+      "order": 0,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "AUDIO",
+          "type": "AUDIO",
+          "links": [
+            51
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.3.32",
+        "Node name for S&R": "LoadAudio"
+      },
+      "widgets_values": [
+        "4_jidaohuman-megatts3_0.flac",
+        null,
+        null
+      ]
+    },
+    {
+      "id": 31,
+      "type": "AVSaveSubtitles",
+      "pos": [
+        819.79541015625,
+        213.02178955078125
+      ],
+      "size": [
+        270,
+        82
+      ],
+      "flags": {
+        "collapsed": true
+      },
+      "order": 3,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "subtitles",
+          "type": "STRING",
+          "widget": {
+            "name": "subtitles"
+          },
+          "link": 45
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "Node name for S&R": "AVSaveSubtitles"
+      },
+      "widgets_values": [
+        "",
+        "subtitles"
+      ]
+    },
+    {
+      "id": 34,
+      "type": "AVASRTimestamp",
+      "pos": [
+        467.7745361328125,
+        206.14337158203125
+      ],
+      "size": [
+        270,
+        122
+      ],
+      "flags": {},
+      "order": 1,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "audio",
+          "type": "AUDIO",
+          "link": 51
+        }
+      ],
+      "outputs": [
+        {
+          "name": "text",
+          "type": "STRING",
+          "links": null
+        },
+        {
+          "name": "json_result",
+          "type": "STRING",
+          "links": null
+        },
+        {
+          "name": "asr_result",
+          "type": "ASRRESULT",
+          "links": [
+            52
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "AVASRTimestamp"
+      },
+      "widgets_values": [
+        300,
+        false
+      ]
+    },
+    {
+      "id": 30,
+      "type": "AVFormat2Subtitle",
+      "pos": [
+        485.57000732421875,
+        415.3599853515625
+      ],
+      "size": [
+        210,
+        88
+      ],
+      "flags": {
+        "collapsed": false
+      },
+      "order": 2,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "asr_result",
+          "type": "ASRRESULT",
+          "link": 52
+        }
+      ],
+      "outputs": [
+        {
+          "name": "subtitle",
+          "type": "STRING",
+          "links": [
+            45,
+            53
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "AVFormat2Subtitle"
+      },
+      "widgets_values": [
+        ""
+      ]
+    },
+    {
+      "id": 35,
+      "type": "Bjornulf_ShowText",
+      "pos": [
+        804.8345336914062,
+        269.0433654785156
+      ],
+      "size": [
+        398,
+        387
+      ],
+      "flags": {},
+      "order": 4,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "text_value",
+          "type": "STRING",
+          "link": 53
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "cnr_id": "bjornulf_custom_nodes",
+        "ver": "1.0.0",
+        "Node name for S&R": "Bjornulf_ShowText"
+      },
+      "widgets_values": [
+        "1\n00:00:00,150 --> 00:00:02,610\n直观来说测试时试错的步骤只有\n\n2\n00:00:02,610 --> 00:00:05,110\n当能带来信息增益的情况下才有益\n\n3\n00:00:05,410 --> 00:00:07,130\n然而人们并没有在all训练中\n\n4\n00:00:07,130 --> 00:00:10,225\n告诉模型试错和反思带来的信息增益\n\n"
+      ]
+    }
+  ],
+  "links": [
+    [
+      45,
+      30,
+      0,
+      31,
+      0,
+      "STRING"
+    ],
+    [
+      51,
+      2,
+      0,
+      34,
+      0,
+      "AUDIO"
+    ],
+    [
+      52,
+      34,
+      2,
+      30,
+      0,
+      "ASRRESULT"
+    ],
+    [
+      53,
+      30,
+      0,
+      35,
+      0,
+      "STRING"
+    ]
+  ],
+  "groups": [],
+  "config": {},
+  "extra": {
+    "frontendVersion": "1.18.9",
+    "VHS_latentpreview": false,
+    "VHS_latentpreviewrate": 0,
+    "VHS_MetadataImage": true,
+    "VHS_KeepIntermediate": true
+  },
+  "version": 0.4
+}
@@ -0,0 +1,136 @@
+import jieba
+import string
+import json
+import Levenshtein
+import re
+
+class Format2Subtitle:
+    def __init__(self, asr_result, ori_text):
+        self.asr_result = asr_result
+        self.ori_text = ori_text  
+    def ms_to_srt_time(self, ms):
+        """
+        将毫秒转换为SRT时间格式 (HH:MM:SS,mmm)
+        例如：150毫秒 -> 00:00:00,150
+        """
+        hours = ms // 3600000  # 1小时 = 3,600,000毫秒
+        ms %= 3600000
+        minutes = ms // 60000  # 1分钟 = 60,000毫秒
+        ms %= 60000
+        seconds = ms // 1000   # 1秒 = 1,000毫秒
+        milliseconds = ms % 1000
+        return f"{hours:02d}:{minutes:02d}:{seconds:02d},{milliseconds:03d}"
+    # 查询标点符号位置
+    def find_punctuation_indices(self, words):
+        punctuation = """，。；！？!,;.?:"""
+        for i, word in enumerate(words):
+            if word in punctuation:
+                return i
+        return -1
+
+    # 清除句子中的断句符号
+    def remove_sentence_punctuation(self, sentence):
+        """去除句子中的断句符号"""
+        punctuation = """，。；！？!,;.?:()（）"""
+        for p in punctuation:
+            sentence = sentence.replace(p, ' ')
+        sentence = sentence.strip()
+        return sentence
+
+    def align_char_timestamps(self, original_sentence, recognized_text, word_timestamps, text_length=8):
+        """
+        将词语级时间戳对齐到字符级时间戳
+        
+        参数:
+            original_sentence (str): 带标点的原始句子
+            recognized_text (str): 空格分隔的无标点识别文本
+            word_timestamps (list): 词语级时间戳列表 [[start1, end1], [start2, end2], ...]
+            
+        返回:
+            list: 原始字符级时间戳列表 [[char_start, char_end], ...]
+        """
+        if not original_sentence:
+            original_sentence = "".join(recognized_text.split())
+
+        words = jieba.lcut(original_sentence)
+
+        # print(len(original_sentence), len(word_timestamps), len(recognized_text))
+        recognized_words = recognized_text.split()
+
+        # 4. 为每个字符分配时间戳
+        char_timestamps = []
+        
+        while True:
+            if len(words) < 10:
+                break
+            # 查询列表中的标点符号的位置
+            punctuation_indices = self.find_punctuation_indices(words)
+            # print("标点符号位置：",punctuation_indices)
+            # 如果位置小于10，怎使用这个位置分割
+            if punctuation_indices > 10 or punctuation_indices <= 2:
+                _words = words[:text_length]
+                words = words[text_length:]
+            else:
+                _words = words[:punctuation_indices]
+                words = words[punctuation_indices:]
+
+            text = ''.join(_words)
+            # 计算词语持续时间
+            best_index = -1
+            best_score = 9999
+            for index in range(text_length*4):
+                _recognized_words = recognized_words[:index]
+                _recognized_text = ''.join(_recognized_words)
+                if text == _recognized_text:
+                    break
+                # 计算编辑距离
+                distance = Levenshtein.distance(text, _recognized_text)
+                if distance < best_score:
+                    best_index = index
+                    best_score = distance
+            _recognized_words = recognized_words[:best_index]
+            recognized_words = recognized_words[best_index:]
+
+            timestamp = word_timestamps[:best_index]
+            word_timestamps = word_timestamps[best_index:]
+            word_start = timestamp[0][0]
+            word_end = timestamp[-1][1]
+            
+            char_timestamps.append([word_start, word_end, self.remove_sentence_punctuation(text)])
+        # 处理最后一段
+        if  len(word_timestamps) > 0:
+            text = ''.join(words)
+            timestamp = word_timestamps
+            word_start = timestamp[0][0]
+            word_end = timestamp[-1][1]
+            char_timestamps.append([word_start, word_end, self.remove_sentence_punctuation(text)])
+        return char_timestamps
+
+    def format_subtitle(self, char_timestamps):
+        """
+        [230, 2210, '一个能够从音频中生成面部']
+        [230, 5430, '运动系数的工具']
+        转为字幕文件
+        """
+        srt_content = ""
+        counter = 1
+        for item in char_timestamps:
+            start_time = self.ms_to_srt_time(item[0])
+            end_time = self.ms_to_srt_time(item[1])
+            text = item[2]
+            srt_content += f"{counter}\n{start_time} --> {end_time}\n{text}\n\n"
+            # print(srt_content)
+            counter += 1
+        return srt_content
+    
+    def pipeline(self, output=None):
+        # print(rec_result)
+        recognized_text = self.asr_result['text']
+        word_timestamps = self.asr_result['timestamp']
+        original_sentence = self.ori_text
+        char_timestamps = self.align_char_timestamps(original_sentence, recognized_text, word_timestamps)
+        subtitle_text = self.format_subtitle(char_timestamps)
+        if output is not None:
+            with open(output, 'w', encoding='utf-8') as f:
+                f.write(subtitle_text)
+        return subtitle_text
@@ -0,0 +1,220 @@
+import folder_paths
+import os
+import comfy.model_management as mm
+import time
+import torchaudio
+import torchvision.utils as vutils
+import torch
+import json
+
+from modelscope.pipelines import pipeline
+from modelscope.utils.constant import Tasks
+from funasr import AutoModel
+from .format import Format2Subtitle
+
+
+
+name_maps_ms = {
+    "paraformer": "iic/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch",
+    "paraformer-zh": "iic/speech_seaco_paraformer_large_asr_nat-zh-cn-16k-common-vocab8404-pytorch",
+    "paraformer-en": "iic/speech_paraformer-large-vad-punc_asr_nat-en-16k-common-vocab10020",
+    "paraformer-en-spk": "iic/speech_paraformer-large-vad-punc_asr_nat-en-16k-common-vocab10020",
+    "paraformer-zh-streaming": "iic/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-online",
+    "fsmn-vad": "iic/speech_fsmn_vad_zh-cn-16k-common-pytorch",
+    "ct-punc": "iic/punc_ct-transformer_cn-en-common-vocab471067-large",
+    "ct-punc-c": "iic/punc_ct-transformer_zh-cn-common-vocab272727-pytorch",
+    "fa-zh": "iic/speech_timestamp_prediction-v1-16k-offline",
+    "cam++": "iic/speech_campplus_sv_zh-cn_16k-common",
+    "Whisper-large-v2": "iic/speech_whisper-large_asr_multilingual",
+    "Whisper-large-v3": "iic/Whisper-large-v3",
+    "Qwen-Audio": "Qwen/Qwen-Audio",
+    "emotion2vec_plus_large": "iic/emotion2vec_plus_large",
+    "emotion2vec_plus_base": "iic/emotion2vec_plus_base",
+    "emotion2vec_plus_seed": "iic/emotion2vec_plus_seed",
+    "Whisper-large-v3-turbo": "iic/Whisper-large-v3-turbo",
+}
+
+class AVSpeechTimestamp:
+    infer_ins_cache = None
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "audio": ("AUDIO",),
+                "text": ("STRING",),
+                "unload_model": ("BOOLEAN", {"default": False}),
+            },
+        }
+
+    RETURN_TYPES = ("STRING", "STRING", "ASRRESULT")
+    RETURN_NAMES = ("text", "json_result", "asr_result")
+    FUNCTION = "infer"
+    CATEGORY = "Aven/AV-FunASR"
+    DESCRIPTION = "get speech timestamp"
+
+    def infer(self, audio, text, unload_model):
+        temp_dir = folder_paths.get_temp_directory()
+        os.makedirs(temp_dir, exist_ok=True)
+
+        if AVSpeechTimestamp.infer_ins_cache is None:
+            model_root = os.path.join(folder_paths.models_dir, "ASR/FunASR")
+            model_dir = os.path.join(model_root, name_maps_ms["fa-zh"])
+            
+            os.makedirs(model_dir, exist_ok=True)
+            AVSpeechTimestamp.infer_ins_cache = pipeline(
+                task=Tasks.speech_timestamp,
+                model=model_dir,
+                model_revision="v2.0.4",
+                output_dir=temp_dir
+            )
+        # save audio
+        audio_save_path = os.path.join(temp_dir, f"{int(time.time())}.wav")
+        torchaudio.save(audio_save_path, audio['waveform'].squeeze(0), audio["sample_rate"])
+
+        rec_result = AVSpeechTimestamp.infer_ins_cache(input=(audio_save_path,text), data_type=("sound", "text"))
+        if rec_result:
+            rec_result = rec_result[0]
+
+        # infer
+        if unload_model:
+            import gc
+            if AVSpeechTimestamp.infer_ins_cache is not None:
+                AVSpeechTimestamp.infer_ins_cache = None
+                gc.collect()
+                torch.cuda.empty_cache()
+                print("AVSpeechTimestamp memory cleanup successful")
+        # jr = json.dumps(rec_result, indent=4)
+        text = rec_result.get("text")
+        jr = json.dumps(rec_result, ensure_ascii=False)
+        print(text, jr)
+        return (text, jr, rec_result)
+    
+
+class AVASRTimestamp:
+    infer_ins_cache = None
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "audio": ("AUDIO",),
+                "batch_size_s": ("INT", {"default": 300, "min": 30, "max": 300, "step": 1}),
+                "unload_model": ("BOOLEAN", {"default": False}),
+            },
+        }
+
+    RETURN_TYPES = ("STRING", "STRING", "ASRRESULT")
+    RETURN_NAMES = ("text", "json_result", "asr_result")
+    FUNCTION = "infer"
+    CATEGORY = "Aven/AV-FunASR"
+    DESCRIPTION = "get speech timestamp"
+
+    def infer(self, audio, batch_size_s, unload_model):
+        temp_dir = folder_paths.get_temp_directory()
+        os.makedirs(temp_dir, exist_ok=True)
+
+        if AVASRTimestamp.infer_ins_cache is None:
+            model_root = os.path.join(folder_paths.models_dir, "ASR/FunASR")
+            model_dir = os.path.join(model_root, name_maps_ms["paraformer-zh"])
+            vad_model = os.path.join(model_root, name_maps_ms["fsmn-vad"])
+            os.makedirs(model_dir, exist_ok=True)
+            device = "cuda" if torch.cuda.is_available() else "cpu"
+            AVASRTimestamp.infer_ins_cache = AutoModel(
+                model=model_dir,
+                vad_model=vad_model,
+                punc_model=None, #"ct-punc",
+                device=device  # GPU加速
+            )
+        # save audio
+        audio_save_path = os.path.join(temp_dir, f"{int(time.time())}.wav")
+        torchaudio.save(audio_save_path, audio['waveform'].squeeze(0), audio["sample_rate"])
+
+        rec_result = AVASRTimestamp.infer_ins_cache.generate(input=audio_save_path, batch_size_s=batch_size_s)
+        if rec_result:
+            rec_result = rec_result[0]
+
+        # infer
+        if unload_model:
+            import gc
+            if AVSpeechTimestamp.infer_ins_cache is not None:
+                AVSpeechTimestamp.infer_ins_cache = None
+                gc.collect()
+                torch.cuda.empty_cache()
+                print("AVSpeechTimestamp memory cleanup successful")
+        # jr = json.dumps(rec_result, indent=4)
+        text = rec_result.get("text")
+        jr = json.dumps(rec_result, ensure_ascii=False)
+        print((text, jr, rec_result))
+        return (text, jr, rec_result)
+    
+
+class AVFormat2Subtitle:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "asr_result": ("ASRRESULT",),
+                "text":  ("STRING", {"default": None}),
+            },
+        }
+
+    RETURN_TYPES = ("STRING", )
+    RETURN_NAMES = ("subtitle", )
+    FUNCTION = "format_subtitle"
+    CATEGORY = "Aven/AV-FunASR"
+    DESCRIPTION = "format asr result to subtitle"
+
+    def format_subtitle(self, asr_result, text=None):
+        f = Format2Subtitle(asr_result, ori_text=text)
+        content = f.pipeline()
+        
+        return (content, )
+    
+
+class AVSaveSubtitles:
+    def __init__(self):
+        self.output_dir = folder_paths.get_output_directory()
+        self.type = "output"
+        self.prefix_append = ""
+        self.compress_level = 4
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "subtitles": ("STRING", {"tooltip": "The subtitles to save."}),
+                "filename_prefix": ("STRING", {"default": "subtitles", "tooltip": "The prefix for the file to save. "})
+            },
+        }
+
+    RETURN_TYPES = ()
+    FUNCTION = "save_subtitles"
+
+    OUTPUT_NODE = True
+
+    CATEGORY = "Aven/AV-FunASR"
+    DESCRIPTION = "Saves the subtitles to a file."
+
+    def save_subtitles(self, subtitles, filename_prefix="subtitles"):
+        filename = f"{filename_prefix}-{time.time()}.srt"
+        save_path = os.path.join(self.output_dir, filename)
+        with open(save_path, "w") as f:
+            f.write(subtitles)
+
+        return { "ui": { "text": save_path } }
+
+
+    
+
+NODE_CLASS_MAPPINGS = {
+    "AVSpeechTimestamp": AVSpeechTimestamp,
+    "AVASRTimestamp": AVASRTimestamp,
+    "AVFormat2Subtitle": AVFormat2Subtitle,
+    "AVSaveSubtitles": AVSaveSubtitles,
+}
+
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "AVSpeechTimestamp": "AV Speech Timestamp",
+    "AVASRTimestamp": "AV ASR Timestamp",
+    "AVFormat2Subtitle": "AV Format to Subtitle",
+    "AVSaveSubtitles": "AV Save Subtitles",
+}
@@ -0,0 +1,15 @@
+[project]
+name = "ComfyUI-AV-FunASR"
+description = "FunASR wrapper for ComfyUI"
+version = "1.0.0"
+license = {file = "LICENSE"}
+dependencies = ["modelscope", "sortedcontainers", "simplejson", "datasets==2.16.0", "Levenshtein", "soundfile"]
+
+[project.urls]
+Repository = "https://github.com/avenstack/ComfyUI-AV-FunASR"
+#  Used by Comfy Registry https://comfyregistry.org
+
+[tool.comfy]
+PublisherId = "av"
+DisplayName = "ComfyUI-AV-FunASR"
+Icon = ""
@@ -0,0 +1,5 @@
+sortedcontainers
+simplejson
+datasets==2.16.0
+Levenshtein
+modelscope