Merge pull request #348 from shadowcz007/whisper-sensevoice

Whisper sensevoice

__init__.py

@@ -1006,7 +1006,7 @@ from .nodes.ImageNode import DepthViewer_,ImageBatchToList_,ImageListToBatch_,Co
 # from .nodes.Vae import VAELoader,VAEDecode
 from .nodes.ScreenShareNode import ScreenShareNode,FloatingVideo
 
-from .nodes.Audio import AudioPlayNode,SpeechRecognition,SpeechSynthesis
+from .nodes.Audio import AudioPlayNode,SpeechRecognition,SpeechSynthesis,AnalyzeAudioNone
 from .nodes.Utils import CreateJsonNode,KeyInput,IncrementingListNode,ListSplit,CreateLoraNames,CreateSampler_names,CreateCkptNames,CreateSeedNode,TESTNODE_,TESTNODE_TOKEN,AppInfo,IntNumber,FloatSlider,TextInput,ColorInput,FontInput,TextToNumber,DynamicDelayProcessor,LimitNumber,SwitchByIndex,MultiplicationNode
 from .nodes.Mask import PreviewMask_,MaskListReplace,MaskListMerge,OutlineMask,FeatheredMask
 
@@ -1103,6 +1103,7 @@ NODE_CLASS_MAPPINGS = {
     "SpeechRecognition":SpeechRecognition,
     "SpeechSynthesis":SpeechSynthesis,
     "AudioPlay":AudioPlayNode,
+    "AnalyzeAudio":AnalyzeAudioNone,
 
     # Text
     "TextToNumber":TextToNumber,
@@ -1220,6 +1221,7 @@ NODE_DISPLAY_NAME_MAPPINGS = {
     "SpeechSynthesis":"SpeechSynthesis ♾️Mixlab",
     "SpeechRecognition":"SpeechRecognition ♾️Mixlab",
     "AudioPlay":"Preview Audio ♾️Mixlab",
+    "AnalyzeAudio":"Analyze Audio ♾️Mixlab",
 
     # Utils
     "DynamicDelayProcessor":"DynamicDelayByText ♾️Mixlab",
@@ -1433,11 +1435,20 @@ try:
     from .nodes.SenseVoice import SenseVoiceNode
     logging.info('SenseVoice.available')
     NODE_CLASS_MAPPINGS['SenseVoiceNode']=SenseVoiceNode
-    NODE_DISPLAY_NAME_MAPPINGS["SenseVoiceNode"]= "Sense Voice"
+    NODE_DISPLAY_NAME_MAPPINGS["SenseVoiceNode"]= "Sense Voice ♾️Mixlab"
 
 except Exception as e:
     logging.info('SenseVoice.available False' )  
 
+try:
+    from .nodes.Whisper import LoadWhisperModel,WhisperTranscribe
+    logging.info('Whisper.available')
+    NODE_CLASS_MAPPINGS['LoadWhisperModel_']=LoadWhisperModel
+    NODE_CLASS_MAPPINGS['WhisperTranscribe_']=WhisperTranscribe
+    NODE_DISPLAY_NAME_MAPPINGS["LoadWhisperModel_"]= "Load Whisper Model ♾️Mixlab"
+    NODE_DISPLAY_NAME_MAPPINGS["WhisperTranscribe_"]= "Whisper Transcribe ♾️Mixlab"
 
+except Exception as e:
+    logging.info('Whisper.available False' )  
 
 logging.info('\033[93m -------------- \033[0m')

nodes/Audio.py

@@ -3,6 +3,101 @@ import os
 import folder_paths
 import torchaudio
 
+class AnyType(str):
+  """A special class that is always equal in not equal comparisons. Credit to pythongosssss"""
+
+  def __ne__(self, __value: object) -> bool:
+    return False
+
+any_type = AnyType("*")
+
+
+def analyze_audio_data(audio_data):
+    total_duration = 0
+    total_gap_duration = 0
+    emotion_counts = {}
+    audio_types = set()
+    languages = set()
+
+    for i, entry in enumerate(audio_data):
+        # Calculate the duration of each audio segment
+        start_time = entry['start_time']
+        end_time = entry['end_time']
+        duration = end_time - start_time
+        total_duration += duration
+
+        # Count the emotions
+        if "emotion" in entry:
+            emotion = entry['emotion']
+            if emotion in emotion_counts:
+                emotion_counts[emotion] += 1
+            else:
+                emotion_counts[emotion] = 1
+
+        # Collect the audio types
+        if "audio_type" in entry:
+            audio_types.add(entry['audio_type'])
+
+        if "language" in entry:
+            languages.add(entry['language'])
+
+        # Calculate gap duration if not the last entry
+        if i < len(audio_data) - 1:
+            next_start_time = audio_data[i + 1]['start_time']
+            gap_duration = next_start_time - end_time
+            if gap_duration > 0:
+                total_gap_duration += gap_duration
+
+    # Get the most frequent emotion
+    if len(emotion_counts.keys())>0:
+        most_frequent_emotion = max(emotion_counts, key=emotion_counts.get)
+    else:
+        most_frequent_emotion=None
+
+    # Convert audio_types set to list for better readability
+    audio_types = list(audio_types)
+
+    languages=list(languages)
+
+    # Print the results
+    print(f"Total Effective Duration: {total_duration:.2f} seconds")
+    print(f"Total Gap Duration: {total_gap_duration:.2f} seconds")
+    print(f"Emotion Changes: {emotion_counts}")
+    print(f"Most Frequent Emotion: {most_frequent_emotion}")
+    print(f"Audio Types: {audio_types}")
+
+
+    return {
+        "total_duration": total_duration,
+        "total_gap_duration": total_gap_duration,
+        "emotion_changes": emotion_counts,
+        "most_frequent_emotion": most_frequent_emotion,
+        "audio_types": audio_types,
+        "languages":languages
+    }
+
+
+# 分析音频数据
+class AnalyzeAudioNone:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { 
+                    "json":(any_type,),}, 
+                }
+    
+    RETURN_TYPES = (any_type,)
+    RETURN_NAMES = ("result",)
+
+    FUNCTION = "run"
+
+    CATEGORY = "♾️Mixlab/Audio"
+
+    def run(self,json):
+        result=analyze_audio_data(json)
+        return (result,)
+
+
+
 class SpeechRecognition:
     @classmethod
     def INPUT_TYPES(s):

nodes/ChatGPT.py

@@ -786,9 +786,12 @@ class JsonRepair:
     def INPUT_TYPES(s):
         return {
             "required": {
-                 "json_string":("STRING", {"forceInput": True,}), 
-                 "key":("STRING", {"multiline": False,"dynamicPrompts": False,"default": ""}),
-            }
+                "json_string":("STRING", {"forceInput": True,}), 
+                "key":("STRING", {"multiline": False,"dynamicPrompts": False,"default": ""}),
+            },
+            "optional":{
+                "json_string2":("STRING", {"forceInput": True,})
+            },
         }
 
     INPUT_IS_LIST = False
@@ -800,8 +803,11 @@ class JsonRepair:
 
     CATEGORY = "♾️Mixlab/GPT"
 
-    def run(self, json_string,key=""):
+    def run(self, json_string,key="",json_string2=None):
 
+        if not isinstance(json_string, str):
+            json_string=json.dumps(json_string)
+    
         json_string=extract_json_strings(json_string)
         # print(json_string)
         good_json_string = repair_json(json_string)
@@ -809,6 +815,20 @@ class JsonRepair:
         # 将 JSON 字符串解析为 Python 对象
         data = json.loads(good_json_string)
 
+        if json_string2!=None:
+            if not isinstance(json_string2, str):
+                json_string2=json.dumps(json_string2)
+    
+            json_string2=extract_json_strings(json_string2)
+            # print(json_string)
+            good_json_string2 = repair_json(json_string2)
+
+            # 将 JSON 字符串解析为 Python 对象
+            data2 = json.loads(good_json_string2)
+
+            data={**data, **data2}
+
+
         v=""
         if key!="" and (key in data):
             v=data[key]

nodes/MiniCPMNode.py

@@ -1,4 +1,5 @@
 # Referenced some code：https://github.com/IuvenisSapiens/ComfyUI_MiniCPM-V-2_6-int4
+# https://github.com/CY-CHENYUE/ComfyUI-MiniCPM-Plus
 
 import os
 import torch
@@ -35,6 +36,7 @@ class MiniCPM_VQA_Simple:
                 "images": ("IMAGE",),
                 "text": ("STRING", {"default": "", "multiline": True}),   
                 "seed": ("INT", {"default": -1}),  # add seed parameter, default is -1
+                "extract_keywords":("BOOLEAN", {"default": False}),
                 "temperature": (
                     "FLOAT",
                     {
@@ -46,7 +48,9 @@ class MiniCPM_VQA_Simple:
             
         }
 
-    RETURN_TYPES = ("STRING",)
+    RETURN_TYPES = ("STRING","STRING",)
+    RETURN_NAMES = ("result","keywords",)
+
     FUNCTION = "inference"
     CATEGORY = "♾️Mixlab/Image"
 
@@ -55,6 +59,7 @@ class MiniCPM_VQA_Simple:
         images,
         text, 
         seed,  # add seed parameter, default is -1
+        extract_keywords,
         temperature,
         keep_model_loaded,
     ):
@@ -90,6 +95,7 @@ class MiniCPM_VQA_Simple:
                 torch_dtype=torch.bfloat16 if self.bf16_support else torch.float16,
             )
 
+
         with torch.no_grad():
             images = images.permute([0, 3, 1, 2])
             images = [ToPILImage()(img).convert("RGB") for img in images]
@@ -113,6 +119,30 @@ class MiniCPM_VQA_Simple:
                 # max_new_tokens=max_new_tokens,
                 **params,
             )
+
+            keyword_result=""
+
+            if extract_keywords:#extract_keywords
+                keyword_prompt = f"""Please extract keywords from the following text, including all occurrences of language (e.g. Chinese, English, etc.)：
+                [[[{result}]]]
+                Please list the keywords extracted, separated by commas. Make sure to include all important words, no matter what language. For English words, please keep the original case."""
+
+                keyword_msgs = [{'role': 'user', 'content': keyword_prompt}]
+                keyword_result = self.model.chat(
+                    image=None,
+                    msgs=keyword_msgs,
+                    tokenizer=self.tokenizer,
+                    sampling=True,
+                    # top_k=top_k,
+                    # top_p=top_p,
+                    temperature=temperature,
+                    # repetition_penalty=repetition_penalty,
+                    # max_new_tokens=max_new_tokens,
+                    **params,
+                )
+                print("keyword_result",keyword_result)
+
+               
             # offload model to GPU
             # self.model = self.model.to(torch.device("cpu"))
             # self.model.eval()
@@ -124,4 +154,4 @@ class MiniCPM_VQA_Simple:
                 torch.cuda.empty_cache()  # release GPU memory
                 torch.cuda.ipc_collect()
             # print(result)
-            return (result,)
+            return (result,keyword_result,)

nodes/SenseVoice.py

@@ -48,6 +48,9 @@ def format_to_srt(channel_id, start_time_ms, end_time_ms, asr_result):
 
     pattern = r"<\|(.+?)\|><\|(.+?)\|><\|(.+?)\|><\|(.+?)\|>(.+)"
     match = re.match(pattern,asr_result)
+    print('#format_to_srt',match,asr_result)
+    if match==None:
+        return None, None, None, None,None,start_time,end_time,None
     lang, emotion, audio_type, itn, text = match.groups()
      # 😊 表示高兴，😡 表示愤怒，😔 表示悲伤。对于音频事件，🎼 表示音乐，😀 表示笑声，👏 表示掌声
     
@@ -115,16 +118,17 @@ class SenseVoiceProcessor:
                     part[1], 
                     asr_result)
 
-                results.append({
-                    "language":lang, 
-                    "emotion":emotion,
-                    "audio_type":audio_type, 
-                    "itn":itn,
-                    "srt_content":srt_content,
-                    "start_time":start_time,
-                    "end_time":end_time,
-                    "text":text
-                })
+                if lang!=None:
+                    results.append({
+                        "language":lang, 
+                        "emotion":emotion,
+                        "audio_type":audio_type, 
+                        "itn":itn,
+                        "srt_content":srt_content,
+                        "start_time":start_time,
+                        "end_time":end_time,
+                        "text":text
+                    })
 
             self.vad.vad.all_reset_detection()
             pbar.update(1)  # 更新进度条
@@ -168,8 +172,9 @@ class SenseVoiceNode:
 
     OUTPUT_NODE = True
     FUNCTION = "run" 
-    RETURN_TYPES = (any_type,)
-    RETURN_NAMES = ("result",)
+    
+    RETURN_TYPES = (any_type,"STRING","STRING","FLOAT",)
+    RETURN_NAMES = ("result","srt","text","total_seconds",)
 
     def run(self,audio,device,language,num_threads,use_int8,use_itn ):
  
@@ -200,16 +205,22 @@ class SenseVoiceNode:
 
         if 'waveform' in audio and 'sample_rate' in audio:
             waveform = audio['waveform']
+            sample_rate = audio['sample_rate']
             # print("Original shape:", waveform.shape)  # 打印原始形状
             if waveform.ndim == 3 and waveform.shape[0] == 1:  # 检查是否为三维且 batch_size 为 1
                 waveform = waveform.squeeze(0)  # 移除 batch_size 维度
-                waveform_numpy = waveform.numpy().transpose(1, 0)  # 转换为 (num_samples, num_channels)
             else:
                 raise ValueError("Unexpected waveform dimensions")
 
-            _sample_rate = audio['sample_rate']
+            print("waveform.shape:", waveform.shape) 
+            total_length_seconds = waveform.shape[1] / sample_rate
 
-        results=self.processor.process_audio(waveform_numpy, _sample_rate, language, use_itn)
+            waveform_numpy = waveform.numpy().transpose(1, 0)  # 转换为 (num_samples, num_channels)
 
+        results=self.processor.process_audio(waveform_numpy, sample_rate, language, use_itn)
 
-        return (results,)
+        srt_content="\n".join([s['srt_content'] for s in results])
+        text="\n".join([s['text'] for s in results])
+
+        return (results,srt_content,text,total_length_seconds,)
+ 

nodes/Whisper.py

@@ -0,0 +1,173 @@
+import os,re
+import sys,time
+from pathlib import Path
+import torchaudio
+import hashlib
+import torch
+import folder_paths
+import comfy.utils
+
+from faster_whisper import WhisperModel
+
+class AnyType(str):
+  """A special class that is always equal in not equal comparisons. Credit to pythongosssss"""
+
+  def __ne__(self, __value: object) -> bool:
+    return False
+
+any_type = AnyType("*")
+
+def get_model_dir(m):
+    try:
+        return folder_paths.get_folder_paths(m)[0]
+    except:
+        return os.path.join(folder_paths.models_dir, m)
+
+
+
+whisper_model_path=get_model_dir('whisper')
+
+model_sizes=[
+                d for d in os.listdir(whisper_model_path) if os.path.isdir(
+                    os.path.join(whisper_model_path, d)
+                    ) and os.path.isfile(os.path.join(os.path.join(whisper_model_path, d), "config.json"))
+                    ]
+
+
+class LoadWhisperModel:
+    def __init__(self):
+        self.model = None
+        self.device="cuda" if torch.cuda.is_available() else "cpu"
+        self.model_size=model_sizes[0]
+        self.compute_type='float16'
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {
+            "model_size": (model_sizes,),
+            "device": (["auto","cpu"],),
+            "compute_type": (["float16","int8_float16","int8"],),
+                             },
+                }
+    
+    RETURN_TYPES = ("WHISPER",)
+    RETURN_NAMES = ("whisper_model",)
+
+    FUNCTION = "run"
+
+    CATEGORY = "♾️Mixlab/Audio/Whisper"
+
+    INPUT_IS_LIST = False
+    OUTPUT_IS_LIST = (False,)
+
+    def run(self,model_size,device,compute_type):
+
+        if device=="auto" and self.device!='cuda':
+            self.device="cuda" if torch.cuda.is_available() else "cpu"
+            self.model=None
+
+        if device=='cpu' and self.device!='cpu':
+            self.device="cpu"
+            self.model=None
+        
+        if model_size!= self.model_size:
+            self.model_size=model_size
+            self.model=None
+
+        if compute_type!=self.compute_type:
+            self.compute_type=compute_type
+            self.model=None
+
+        if self.model==None:
+            self.model = WhisperModel(
+                os.path.join(whisper_model_path, self.model_size), 
+                device=self.device,
+                compute_type=self.compute_type
+                )
+
+        return (self.model,)
+    
+
+class WhisperTranscribe:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {
+                                "whisper_model": ("WHISPER",),
+                                "audio": ("AUDIO",),
+                             },
+                }
+    
+    RETURN_TYPES = (any_type,"STRING","STRING","FLOAT",)
+    RETURN_NAMES = ("result","srt","text","total_seconds",)
+
+    FUNCTION = "run"
+
+    CATEGORY = "♾️Mixlab/Audio/Whisper"
+
+    INPUT_IS_LIST = False
+    # OUTPUT_IS_LIST = (False,False,False,)
+    
+    def run(self,whisper_model,audio):
+
+        if 'audio_path' in audio and (not 'waveform' in audio):
+            waveform, sample_rate = torchaudio.load(audio['audio_path'])
+            waveform=waveform.mean(0)
+            total_length_seconds = waveform.shape[0] / sample_rate
+            waveform=waveform.numpy()
+            
+        elif 'waveform' in audio and 'sample_rate' in audio:
+            print("Original shape:", audio["waveform"].shape, isinstance(audio["waveform"], torch.Tensor))  # 打印原始形状
+            waveform = audio["waveform"].squeeze(0)  # Remove the added batch dimension
+            sample_rate = audio["sample_rate"]
+
+            # if audio_sf != sampling_rate:
+            #     waveform = torchaudio.functional.resample(
+            #         waveform, orig_freq=audio_sf, new_freq=sampling_rate
+            #     )
+            
+            waveform=waveform.mean(0)
+
+            total_length_seconds = waveform.shape[0] / sample_rate
+
+            waveform=waveform.numpy() #whisper_model.transcribe 旧版不支持直接传tensor，先用numpy
+        
+        segments, info = whisper_model.transcribe(waveform, beam_size=5)
+
+        print("Detected language '%s' with probability %f" % (info.language, info.language_probability))
+
+        # Function to format time for SRT
+        def format_time(seconds):
+            millis = int((seconds - int(seconds)) * 1000)
+            hours, remainder = divmod(int(seconds), 3600)
+            minutes, seconds = divmod(remainder, 60)
+            return f"{hours:02}:{minutes:02}:{seconds:02},{millis:03}"
+
+        # Prepare SRT content as a string
+        results = []
+        for i, segment in enumerate(segments):
+            start_time = format_time(segment.start)
+            end_time = format_time(segment.end)
+            srt_content = f"{i + 1}\n"
+            srt_content += f"{start_time} --> {end_time}\n"
+
+            text=segment.text.strip()
+
+            srt_content += f"{text}\n\n"
+
+            start_time=segment.start
+            end_time=segment.end
+            
+
+            results.append({ 
+                    "srt_content":srt_content,
+                    "start_time":start_time,
+                    "end_time":end_time,
+                    "text":text,
+                    "language":[info.language]
+            })
+        
+        srt_content="\n".join([s['srt_content'] for s in results])
+        text="\n".join([s['text'] for s in results])
+
+        return (results,srt_content,text,total_length_seconds,)
+

pyproject.toml

@@ -1,7 +1,7 @@
 [project]
 name = "comfyui-mixlab-nodes"
 description = "3D, ScreenShareNode & FloatingVideoNode, SpeechRecognition & SpeechSynthesis, GPT, LoadImagesFromLocal, Layers, Other Nodes, ..."
-version = "0.43.0"
+version = "0.44.0"
 license = "MIT"
 dependencies = ["numpy", "pyOpenSSL", "watchdog", "opencv-python-headless", "matplotlib", "openai", "simple-lama-inpainting", "clip-interrogator==0.6.0", "transformers>=4.36.0", "lark-parser", "imageio-ffmpeg", "rembg[gpu]", "omegaconf==2.3.0", "Pillow>=9.5.0", "einops==0.7.0", "trimesh>=4.0.5", "huggingface-hub", "scikit-image"]
 

requirements.txt

@@ -27,4 +27,9 @@ scenedetect[opencv-headless]
 hydra-core>=1.3.2
 loralib>=0.1.2
 natsort>=8.4.0
-# simple-lama-inpainting
+
+#simple-lama-inpainting
+
+git+https://github.com/shadowcz007/SenseVoice-python.git
+
+faster_whisper

web/javascript/checkVersion_mixlab.js

@@ -3,7 +3,7 @@ import { app } from '../../../scripts/app.js'
 const repoOwner = 'shadowcz007' // 替换为仓库的所有者
 const repoName = 'comfyui-mixlab-nodes' // 替换为仓库的名称
 
-const version = 'v0.43.0'
+const version = 'v0.44.0'
 
 fetch(`https://api.github.com/repos/${repoOwner}/${repoName}/releases/latest`)
   .then(response => response.json())

web/javascript/ui_mixlab.js

@@ -544,7 +544,7 @@ async function getCustomnodeMappings () {
     const data = (await get_nodes_map()).data
     window._nodes_maps = data
   }
-  console.log('#getCustomnodeMappings', window._nodes_maps)
+  // console.log('#getCustomnodeMappings', window._nodes_maps)
   for (let url in window._nodes_maps) {
     let n = window._nodes_maps[url]
     for (let node of n[0]) {