import torch from transformers import pipeline from transformers.pipelines.audio_utils import ffmpeg_read import gradio as gr import os hugapikey=os.environ['openaikey'] genaikey=os.environ['genaikey'] MODEL_NAME = "seiching/whisper-small-seiching" #MODEL_NAME = "openai/whisper-small" BATCH_SIZE = 8 device = 0 if torch.cuda.is_available() else "cpu" pipe = pipeline( task="automatic-speech-recognition", model=MODEL_NAME, chunk_length_s=30, device=device, ) from openai import OpenAI from concurrent.futures import ThreadPoolExecutor import tiktoken usemodelname='gpt-4-0125-preview' def call_openai_api(openaiobj,transcription): response = openaiobj.chat.completions.create( #model="gpt-3.5-turbo", model=usemodelname, temperature=0, messages=[ { "role": "system", "content": "你是專業的會議紀錄製作員,請根據由語音辨識軟體將會議錄音所轉錄的逐字稿，也請注意逐字稿可能有錯,討論內容細節請略過,請列出經主席確認的會議決議，並要用比較正式及容易閱讀的寫法，避免口語化" }, { "role": "user", "content": transcription } ] ) return response.choices[0].message.content def call_openai_summary(openaiobj,transcription): response = openaiobj.chat.completions.create( #model="gpt-3.5-turbo", model=usemodelname, temperature=0, messages=[ { "role": "system", "content": "你是專業的會議紀錄製作員，請根據分段的會議決證，彙整成正式會議紀錄" }, { "role": "user", "content": transcription } ] ) return response.choices[0].message.content def split_into_chunks(text, tokens=127900): #encoding = tiktoken.encoding_for_model('gpt-3.5-turbo') encoding = tiktoken.encoding_for_model(usemodelname) words = encoding.encode(text) chunks = [] for i in range(0, len(words), tokens): chunks.append(' '.join(encoding.decode(words[i:i + tokens]))) return chunks def process_chunks(openaikeystr,inputtext): # openaiobj = OpenAI( # # This is the default and can be omitted # api_key=openaikeystr, # ) if hugapikey=='test': realkey=openaikeystr else: realkey=hugapikey #openaiojb =OpenAI(base_url="http://localhost:1234/v1", api_key="not-needed") openaiobj =OpenAI( api_key=realkey) text = inputtext #openaikey.set_key(openaikeystr) #print('process_chunk',openaikey.get_key()) chunks = split_into_chunks(text) response='這是分段會議紀錄結果\n\n' i=1 for chunk in chunks: response=response+'第' +str(i)+'段\n'+call_openai_api(openaiobj,chunk)+'\n\n' i=i+1 # response=response+call_openai_summary(openaiobj,chunk) if i>2: finalresponse=response+'\n\n 這是根據以上分段會議紀錄彙編如下 \n\n' +call_openai_api(openaiobj,response) else: finalresponse=response return finalresponse # # Processes chunks in parallel # with ThreadPoolExecutor() as executor: # responses = list(executor.map(call_openai_api, [openaiobj,chunks])) # return responses import torch from transformers import pipeline from transformers.pipelines.audio_utils import ffmpeg_read import gradio as gr MODEL_NAME = "seiching/whisper-small-seiching" BATCH_SIZE = 8 transcribe_text="this is a test" device = 0 if torch.cuda.is_available() else "cpu" pipe = pipeline( task="automatic-speech-recognition", model=MODEL_NAME, chunk_length_s=30, device=device, ) # Copied from https://github.com/openai/whisper/blob/c09a7ae299c4c34c5839a76380ae407e7d785914/whisper/utils.py#L50 def format_timestamp(seconds: float, always_include_hours: bool = False, decimal_marker: str = "."): if seconds is not None: milliseconds = round(seconds * 1000.0) hours = milliseconds // 3_600_000 milliseconds -= hours * 3_600_000 minutes = milliseconds // 60_000 milliseconds -= minutes * 60_000 seconds = milliseconds // 1_000 milliseconds -= seconds * 1_000 hours_marker = f"{hours:02d}:" if always_include_hours or hours > 0 else "" return f"{hours_marker}{minutes:02d}:{seconds:02d}{decimal_marker}{milliseconds:03d}" else: # we have a malformed timestamp so just return it as is return seconds def transcribe(file, task, return_timestamps): outputs = pipe(file, batch_size=BATCH_SIZE, generate_kwargs={"task": task,"language": "chinese",}, return_timestamps=return_timestamps) text = outputs["text"] if return_timestamps: timestamps = outputs["chunks"] timestamps = [ f"[{format_timestamp(chunk['timestamp'][0])} -> {format_timestamp(chunk['timestamp'][1])}] {chunk['text']}" for chunk in timestamps ] text = "\n".join(str(feature) for feature in timestamps) global transcribe_text transcribe_text=text # with open('asr_resul.txt', 'w') as f: # f.write(text) # ainotes=process_chunks(text) # with open("ainotes_result.txt", "a") as f: # f.write(ainotes) return text demo = gr.Blocks() mic_transcribe = gr.Interface( fn=transcribe, inputs=[ gr.inputs.Audio(source="microphone", type="filepath", optional=True), gr.inputs.Radio(["transcribe", "translate"], label="Task", default="transcribe"), gr.inputs.Checkbox(default=False, label="Return timestamps"), ], outputs="text", layout="horizontal", theme="huggingface", title="會議紀錄小幫手AINotes", description=( "可由麥克風錄音或上傳語音檔" f" 使用這個模型 [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) 先做語音辨識再做會議紀錄摘要" " 長度沒有限制" ), allow_flagging="never", ) file_transcribe = gr.Interface( fn=transcribe, inputs=[ gr.inputs.Audio(source="upload", optional=True, label="Audio file", type="filepath"), gr.inputs.Radio(["transcribe", "translate"], label="Task", default="transcribe"), gr.inputs.Checkbox(default=False, label="Return timestamps"), ], outputs="text", layout="horizontal", theme="huggingface", title="會議紀錄小幫手AINotes", description=( "可由麥克風錄音或上傳語音檔" f" 使用這個模型 [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) 先做語音辨識再做會議紀錄摘要" " 長度沒有限制" ), # examples=[ # ["./example.flac", "transcribe", False], # ["./example.flac", "transcribe", True], # ], cache_examples=True, allow_flagging="never", ) import google.generativeai as genai def gpt4write(apikeystr,transcribe_text): return 'ok' def gewritenote(inputscript): api_key = genaikey genai.configure(api_key = api_key) model = genai.GenerativeModel('gemini-pro') genprompt='你是專業的會議紀錄製作員,請根據由語音辨識軟體將會議錄音所轉錄的逐字稿，也請注意逐字稿可能有錯,討論內容細節請略過,請列出經主席確認的會議決議，並要用比較正式及容易閱讀的寫法，避免口語化' genprompt=genprompt+'#'+inputscript+'#' response = model.generate_content( genprompt) return response.text def writenotes( LLMmodel,apikeystr,inputscript): #text=transcribe_text #openaikey.set_key(inputkey) #openaikey = OpenAIKeyClass(inputkey) print('ok') if len(inputscript)>10: #有資料表示不是來自語音辨識結果 transcribe_text=inputscript if LLMmodel=="gpt-3.5-turbo": ainotestext=process_chunks(apikeystr,transcribe_text) elif LLMmodel=="gpt-4-0125-preview": ainotestext=gpt4write(apikeystr,transcribe_text) elif LLMmodel=='gemini': ainotestext=gewritenote(inputscript) # ainotestext=inputscript #ainotestext="" # with open('asr_resul.txt', 'w') as f: # #print(transcribe_text) # # f.write(inputkey) # f.write(transcribe_text) # with open('ainotes.txt','w') as f: # f.write(ainotestext) return ainotestext ainotes = gr.Interface( fn=writenotes, inputs=[ gr.inputs.Radio(["gpt-3.5-turbo", "gpt-4-0125-preview","gemini"], label="LLMmodel", default="gpt-3.5-turbo"),gr.Textbox(label="使用GPT請輸入OPEN AI API KEY",placeholder="請輸入sk..."),gr.Textbox(label="逐字稿",placeholder="若沒有做語音辨識，請輸入逐字稿")], outputs="text", layout="horizontal", theme="huggingface", title="會議紀錄小幫手AINotes", description=( "可由麥克風錄音或上傳語音檔，並將本逐字稿欄位清空，若有逐字稿可以直接貼在逐字稿" f" 使用這個模型 [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) 先做語音辨識再做會議紀錄摘要" " 長度沒有限制" ), # examples=[ # ["./example.flac", "transcribe", False], # ["./example.flac", "transcribe", True], # ], cache_examples=True, allow_flagging="never", ) with demo: gr.TabbedInterface([file_transcribe,mic_transcribe,ainotes], ["語音檔辨識","麥克風語音檔辨識","產生會議紀錄" ]) demo.launch(enable_queue=True)