Upload 2 files

asrmodel/app.py

@@ -0,0 +1,170 @@
+# -*- coding: utf-8 -*-
+"""asrdemo.ipynb
+
+Automatically generated by Colab.
+
+Original file is located at
+    https://colab.research.google.com/drive/1A3MDkQEkPUoCpiksvTsAcfVqfNC7-Uxe
+"""
+
+# !pip install gradio transformers datasets torchaudio accelerate unsloth -q
+# !pip install --upgrade protobuf==5.29.1 fsspec==2025.3.2
+# !pip install torchaudio
+# !pip install deep-translator
+# !pip install jiwer
+# !pip install peft unsloth
+
+# from google.colab import drive
+# drive.mount('/content/drive')
+
+# ✨ Step 2: 导入库
+import gradio as gr
+import torch
+import re
+from transformers import SeamlessM4Tv2Model, AutoProcessor
+from unsloth import FastLanguageModel
+
+# ✨ Step 3: 加载 Seamless 模型（用于生成 8-best）
+processor = AutoProcessor.from_pretrained("facebook/seamless-m4t-v2-large")
+seamless_model = SeamlessM4Tv2Model.from_pretrained("facebook/seamless-m4t-v2-large")
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+seamless_model = seamless_model.to(device)
+
+# ✨ Step 4: 加载 DeepSeek 微调模型（后处理）
+deepseek_model, tokenizer = FastLanguageModel.from_pretrained(
+    model_name="unsloth/DeepSeek-R1-Distill-Llama-8B",
+    max_seq_length=2048,
+    dtype=None,
+    load_in_4bit=True
+)
+from unsloth import FastLanguageModel
+from peft import PeftModel
+lora_path = "/content/drive/MyDrive/DeepSeek-FineTune-0422-8bestfinetune/only_lora_adapters_8bestfinetune"
+# 保持 tokenizer 不变
+deepseek_model = PeftModel.from_pretrained(deepseek_model, lora_path)
+deepseek_model.eval()
+# ✨ Step 2: 导入库
+import gradio as gr
+import torch
+import re
+import torchaudio
+from transformers import SeamlessM4Tv2Model, AutoProcessor
+from unsloth import FastLanguageModel
+
+from deep_translator import GoogleTranslator
+from jiwer import wer
+import string
+
+# 初始化翻译函数
+def translate_urdu_to_zh(urdu_text):
+    try:
+        en_text = GoogleTranslator(source='ur', target='en').translate(urdu_text)
+        zh_text = GoogleTranslator(source='en', target='zh-CN').translate(en_text)
+        return en_text.strip(), zh_text.strip()
+    except Exception as e:
+        return "翻译失败", "翻译失败"
+
+# ✨ Step 3: 加载模型（请确保你已定义以下变量）
+# seamless_model, processor, deepseek_model, tokenizer, device
+
+# Prompt 设计
+NEW_PROMPT_STYLE = """
+你是一位精通乌尔都语的专家，负责修正乌尔都语语音转录文本。我给你以下8个结果，请你选出最可能正确的选项，并尝试修改且仅修改可能识别错误的单词，主要考虑发音识别的错误，仅输出最终修改的句子，不要输出其他内容。
+"""
+
+def extract_urdu_sentence(response):
+    urdu_pattern = r'[\u0600-\u06FF\u0750-\u077F\uFB50-\uFDFF\uFE70-\uFEFF\s\.\,\?\!]+'
+    matches = re.findall(urdu_pattern, response)
+    if matches:
+        sentence = re.sub(r'\s+', ' ', matches[-1]).strip()
+        sentence = re.sub(r'^\d+\.\s*', '', sentence)
+        return sentence
+    return response.strip()
+
+def remove_punctuation(text):
+    translator = str.maketrans('', '', string.punctuation)
+    return text.translate(translator).replace("۔", "").replace("،", "").replace("؟", "").strip()
+
+def urdu_asr_postprocess(audio, reference_text):
+    try:
+        sampling_rate, audio_array = audio
+        audio_tensor = torch.tensor(audio_array, dtype=torch.float32)
+
+        if audio_array.dtype == 'int16' or torch.max(torch.abs(audio_tensor)) > 1.0:
+            audio_tensor = audio_tensor / 32768.0
+        if len(audio_tensor.shape) == 1:
+            audio_tensor = audio_tensor.unsqueeze(0)
+        if sampling_rate != 16000:
+            resampler = torchaudio.transforms.Resample(orig_freq=sampling_rate, new_freq=16000)
+            audio_tensor = resampler(audio_tensor)
+            sampling_rate = 16000
+
+        audio_inputs = processor(audios=audio_tensor.squeeze().numpy(), sampling_rate=sampling_rate, return_tensors="pt").to(device)
+        output = seamless_model.generate(
+            **audio_inputs,
+            tgt_lang="__urd__",
+            generate_speech=False,
+            num_beams=8,
+            num_return_sequences=8,
+            pad_token_id=processor.tokenizer.pad_token_id
+        )
+        hypotheses = [processor.tokenizer.decode(seq, skip_special_tokens=True) for seq in output.sequences]
+        input_text = "\n".join([f"{i+1}. {hyp}" for i, hyp in enumerate(hypotheses)])
+        prompt = f"{NEW_PROMPT_STYLE}\n\n输入：{input_text}"
+
+        inputs = tokenizer([prompt], return_tensors="pt").to(deepseek_model.device)
+        outputs = deepseek_model.generate(input_ids=inputs.input_ids, attention_mask=inputs.attention_mask, max_new_tokens=1000)
+        response = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]
+
+        final_output = extract_urdu_sentence(response)
+        final_en, final_zh = translate_urdu_to_zh(final_output)
+
+        # 初始化显示文本
+        final_output_display = f"""🟢 最终转录（乌尔都语）：\n{final_output}"""
+
+        translated_hypotheses = []
+        wer_final = None
+
+        # 判断是否提供参考文本
+        if reference_text and reference_text.strip():
+            ref_clean = remove_punctuation(reference_text)
+            final_clean = remove_punctuation(final_output)
+            wer_final = wer(ref_clean, final_clean)
+            final_output_display += f"\nWER vs 参考文本: {wer_final:.3f}"
+
+        final_output_display += f"""\n\n🌐 翻译（英文）：\n{final_en}\n\n🌐 翻译（中文）：\n{final_zh}\n"""
+
+        for i, hyp in enumerate(hypotheses):
+            en, zh = translate_urdu_to_zh(hyp)
+            wer_i_display = ""
+            if reference_text and reference_text.strip():
+                hyp_clean = remove_punctuation(hyp)
+                wer_i = wer(ref_clean, hyp_clean)
+                wer_i_display = f"\n   WER vs 参考: {wer_i:.3f}"
+            translated_hypotheses.append(f"{i+1}. {hyp}\n   🌐 EN: {en}\n   🌐 中文: {zh}{wer_i_display}\n")
+
+        hypotheses_display = "\n".join(translated_hypotheses)
+        return final_output_display, hypotheses_display, response
+
+    except Exception as e:
+        return f"处理出错: {str(e)}", "", ""
+
+# ✨ Step 7: 构建 Gradio 界面
+demo = gr.Interface(
+    fn=urdu_asr_postprocess,
+    inputs=[
+        gr.Audio(type="numpy", label="上传乌尔都语语音"),
+        gr.Textbox(label="参考文本（可选）", lines=3, placeholder="可选：请输入乌尔都语参考文本以计算 WER")
+    ],
+    outputs=[
+        gr.Textbox(label="最终转录结果"),
+        gr.Textbox(label="8-best 候选结果"),
+        # gr.Textbox(label="DeepSeek 原始输出"),
+    ],
+    title="乌尔都语 ASR + 后处理 Demo",
+    description="上传一段乌尔都语音频，参考文本为可选项。如填写，将计算 WER。",
+    allow_flagging="never"
+)
+
+# ✨ Step 8: 启动 Demo
+demo.launch()

asrmodel/requirements.txt

@@ -0,0 +1,11 @@
+gradio
+transformers
+datasets
+torchaudio
+accelerate
+unsloth
+protobuf==5.29.1
+fsspec==2025.3.2
+deep-translator
+jiwer
+peft