alignment_online.py

import base64
import httpx
import re
import requests
import torch
import torchaudio.functional as F
import torchaudio
import uroman as ur
import logging
import traceback


def convert_to_list_with_punctuation_mixed(text):
    """处理中文文本（可能包含英文单词） - 中文按字符分割，英文单词保持完整"""
    result = []
    text = text.strip()
    
    if not text:
        return result
    
    def is_chinese(char):
        """检查是否是汉字"""
        return '\u4e00' <= char <= '\u9fff'
    
    # 使用更精确的正则表达式来分割文本
    # 匹配：英文单词（含数字）、单个汉字、标点符号
    pattern = r'[a-zA-Z]+[a-zA-Z0-9]*|[\u4e00-\u9fff]|[^\w\s\u4e00-\u9fff]'
    tokens = re.findall(pattern, text)
    
    for token in tokens:
        if not token.strip():  # 跳过空字符
            continue
        
        if re.match(r'^[a-zA-Z]+[a-zA-Z0-9]*$', token):  # 英文单词（可能包含数字）
            result.append(token)
        elif is_chinese(token):  # 单个汉字
            result.append(token)
        else:  # 标点符号等其他字符
            # 标点符号加到前一个词后面
            if result:
                result[-1] += token
            else:
                # 如果是文本开头的标点，单独作为一项
                result.append(token)
    
    return result

def split_and_merge_punctuation(text):
    """处理英文 - 按单词分割，保持单词完整性"""
    # 先按空格拆分文本
    elements = text.split()
    
    # 用于保存最终的结果
    result = []
    
    # 遍历每个拆分后的元素
    for ele in elements:
        # 使用正则表达式提取连续字母、数字和标点
        parts = re.findall(r'[a-zA-Z0-9]+|[^\w\s]+', ele)
        
        # 用于保存拆分后的部分
        merged_parts = []
        
        for i in range(len(parts)):
            if i % 2 == 0:  # 如果是字母或数字部分
                # 将字母或数字部分添加到结果中
                merged_parts.append(parts[i])
            else:  # 如果是标点或其他符号部分
                # 将标点部分与前面的字母或数字部分合并
                if merged_parts:
                    merged_parts[-1] += parts[i]
                else:
                    merged_parts.append(parts[i])
        
        # 将合并后的部分加入最终结果
        result.extend(merged_parts)
    
    return result


def get_aligned_result_text_with_punctuation(alignment_result, text, language):
    """
    将对齐结果转换为正确的文本tokens，英文保持单词级别，中文保持字符级别（但英文单词完整）
    """
    logging.info("start change text to text_tokens")
    
    if language == "EN":
        text_tokens = split_and_merge_punctuation(text)  # 英文按单词分词
    elif language == "ZH":
        text_tokens = convert_to_list_with_punctuation_mixed(text)  # 中文按字符分割，但英文单词保持完整
    else:
        raise ValueError(f"Unsupported language: {language}")

    logging.info(f"Text tokens count: {len(text_tokens)}, Alignment result count: {len(alignment_result)}")
    
    punctuations = set(',.!?;:()[]<>\'\"…·，。；：！？（）【】《》''""\、')

    logging.info("start get align result text with punctuation")
    updated_alignment_result = []
    token_idx = 0
    
    for index, align_item in enumerate(alignment_result):
        if token_idx >= len(text_tokens):
            # 如果text_tokens用完了但还有对齐结果，跳出循环
            logging.warning(f"Text tokens exhausted at index {token_idx}, but alignment has more items")
            break
            
        start = align_item["start"]
        end = align_item["end"]
        text_token = text_tokens[token_idx]
        
        # 检查该 token 后是否有连续标点（仅对中文）
        if language == "ZH":
            while token_idx + 1 < len(text_tokens) and text_tokens[token_idx + 1] in punctuations:
                assert False, "???"  # 这里理论上应该进不去？？
                text_token += text_tokens[token_idx + 1]  # 将标点加入
                token_idx += 1
        else:
            # 英文不需要特殊的标点处理，因为标点已经在split_and_merge_punctuation中处理了
            pass
        
        # 更新对齐结果
        updated_item = {
            "start": start,
            "end": end,
            "transcript": text_token
        }
        updated_item.update({key: align_item[key] for key in align_item if key not in ["start", "end", "transcript"]})
        
        updated_alignment_result.append(updated_item)
        token_idx += 1

    logging.info("end get align result text with punctuation")
    return updated_alignment_result


class AlignmentModel:
    def __init__(self, device, model_dir='/data-mnt/data/wy/X-Codec-2.0/checkpoints'):
        """
        初始化对齐模型并加载必要的资源
        """
        self.device = torch.device(device)
        self.bundle = torchaudio.pipelines.MMS_FA
        model = self.bundle.get_model(with_star=False, dl_kwargs={'model_dir': model_dir}).to(self.device)
        
        # --- 核心优化 ---
        # 使用 torch.compile 对模型进行 JIT 编译
        # mode="max-autotune" 会花费更长时间编译，但能达到最佳性能
        print("Compiling the model... This may take a moment.")
        self.align_model = torch.compile(model, mode="reduce-overhead", fullgraph=True) 
        print("Model compiled successfully.")

        self.uroman = ur.Uroman()
        self.DICTIONARY = self.bundle.get_dict()

    def align(self, emission, tokens):
        """
        执行强对齐
        :param emission: 模型的输出
        :param tokens: 目标 tokens
        :return: 对齐的 tokens 和分数
        """
        alignments, scores = F.forced_align(
            log_probs=emission,
            targets=tokens,
            blank=0
        )
        alignments, scores = alignments[0], scores[0]
        scores = scores.exp()
        return alignments, scores

    def unflatten(self, list_, lengths):
        """
        将一个长列表按照长度拆分成子列表
        :param list_: 长列表
        :param lengths: 各子列表的长度
        :return: 拆分后的子列表
        """
        assert len(list_) == sum(lengths)
        i = 0
        ret = []
        for l in lengths:
            ret.append(list_[i:i + l])
            i += l
        return ret

    def preview_word(self, waveform, spans, num_frames, transcript, sample_rate):
        """
        预览每个单词的开始时间和结束时间
        :param waveform: 音频波形
        :param spans: 单词的跨度
        :param num_frames: 帧数
        :param transcript: 转录文本
        :param sample_rate: 采样率
        :return: 单词的对齐信息
        """
        end = 0
        alignment_result = []
        for span, trans in zip(spans, transcript):
            ratio = waveform.size(1) / num_frames
            x0 = int(ratio * span[0].start)
            x1 = int(ratio * span[-1].end)
            align_info = {
                "transcript": trans,
                "start": round(x0 / sample_rate, 3),
                "end": round(x1 / sample_rate, 3)
            }
            align_info["pause"] = round(align_info["start"] - end, 3)
            align_info["duration"] = round(align_info["end"] - align_info["start"], 3)
            end = align_info["end"]
            alignment_result.append(align_info)
        return alignment_result

    def make_wav_batch(self, wav_list):
        """
        将 wav_list 中的每个 wav 张量填充为相同的长度，返回填充后的张量和每个张量的原始长度。
        :param wav_list: wav 文件列表
        :return: 填充后的音频张量和原始长度
        """
        wav_lengths = torch.tensor([wav.size(0) for wav in wav_list], dtype=torch.long)
        max_length = max(wav_lengths)
        # 确保张量在正确的设备上
        wavs_tensors = torch.zeros(len(wav_list), max_length, device=self.device)
        for i, wav in enumerate(wav_list):
            wav = wav.to(self.device)  # 确保wav在正确的设备上
            wavs_tensors[i, :wav_lengths[i]] = wav
        return wavs_tensors, wav_lengths.to(self.device)

    def get_target(self, transcript, language):
        """
        获取给定转录文本的目标 tokens - 修正版本，保持英文单词完整性
        """
        original_transcript = transcript  # 保存原始文本用于调试
        
        if language == "ZH":
            # 中文处理：保持英文单词完整，只对中文字符进行romanization
            # 使用相同的分词逻辑
            pattern = r'[a-zA-Z]+[a-zA-Z0-9]*|[\u4e00-\u9fff]|[^\w\s\u4e00-\u9fff]'
            tokens = re.findall(pattern, transcript)
            
            # 分别处理中文字符和英文单词
            processed_parts = []
            for token in tokens:
                if not token.strip():
                    continue
                elif re.match(r'^[a-zA-Z]+[a-zA-Z0-9]*$', token):  # 英文单词
                    # 英文单词保持原样，不进行romanization
                    processed_parts.append(token.lower())
                elif '\u4e00' <= token <= '\u9fff':  # 中文字符
                    # 只对中文字符进行romanization
                    romanized = self.uroman.romanize_string(token)
                    processed_parts.append(romanized)
                else:  # 标点符号等
                    # 标点符号直接添加，但会在后续步骤中被过滤掉
                    processed_parts.append(token)
            
            # 用空格连接所有部分
            transcript = ' '.join(processed_parts)
        
        elif language == "EN":
            # 英文处理：保持单词结构，只是清理标点
            pass
        else:
            assert False, f"Unsupported language: {language}"

        # 清理标点符号
        transcript = re.sub(r'[^\w\s]', r' ', transcript)
        TRANSCRIPT = transcript.lower().split()
        
        # 提前获取字典中的特殊符号 token
        star_token = self.DICTIONARY['*']
        tokenized_transcript = []

        # 统一的tokenization逻辑
        for word in TRANSCRIPT:
            # 对每个word中的字符进行token化
            word_tokens = []
            for c in word:
                if c in self.DICTIONARY and c != '-':
                    word_tokens.append(self.DICTIONARY[c])
                else:
                    word_tokens.append(star_token)
            tokenized_transcript.extend(word_tokens)
        
        logging.info(f"Original transcript: {original_transcript}")
        logging.info(f"Processed transcript: {transcript}")
        logging.info(f"Final TRANSCRIPT: {TRANSCRIPT}")
        
        return torch.tensor([tokenized_transcript], dtype=torch.int32, device=self.device)

    def get_alignment_result(self, emission_padded, emission_length, aligned_tokens, alignment_scores, transcript, waveform, language):
        """
        根据给定的 emission 和对齐信息生成对齐结果 - 修正版本
        """
        original_transcript = transcript  # 保存原始文本
        
        if language == "ZH":
            # 使用与get_target相同的处理逻辑
            pattern = r'[a-zA-Z]+[a-zA-Z0-9]*|[\u4e00-\u9fff]|[^\w\s\u4e00-\u9fff]'
            tokens = re.findall(pattern, transcript)
            
            processed_parts = []
            for token in tokens:
                if not token.strip():
                    continue
                elif re.match(r'^[a-zA-Z]+[a-zA-Z0-9]*$', token):  # 英文单词
                    processed_parts.append(token.lower())
                elif '\u4e00' <= token <= '\u9fff':  # 中文字符
                    romanized = self.uroman.romanize_string(token)
                    processed_parts.append(romanized)
                else:  # 标点符号等
                    processed_parts.append(token)
            
            transcript = ' '.join(processed_parts)
        elif language == "EN":
            pass
        else:
            assert False, f"Unsupported language: {language}"
        
        transcript = re.sub(r'[^\w\s]', r' ', transcript)
        emission = emission_padded[:emission_length, :].unsqueeze(0)
        TRANSCRIPT = transcript.lower().split()
        
        token_spans = F.merge_tokens(aligned_tokens, alignment_scores)
        
        # 统一的分组逻辑
        word_spans = self.unflatten(token_spans, [len(word) for word in TRANSCRIPT])
        
        num_frames = emission.size(1)
        
        logging.info(f"Original transcript for alignment: {original_transcript}")
        logging.info(f"Processed TRANSCRIPT: {TRANSCRIPT}")
        
        return self.preview_word(waveform.unsqueeze(0), word_spans, num_frames, TRANSCRIPT, self.bundle.sample_rate)

    def batch_alignment(self, wav_list, transcript_list, language_list):
        """
        批量对齐
        :param wav_list: wav 文件列表
        :param transcript_list: 转录文本列表
        :param language_list: 语言类型列表
        :return: 对齐结果列表
        """
        wavs_tensors, wavs_lengths_tensor = self.make_wav_batch(wav_list)
        logging.info("start alignment model forward")
        with torch.inference_mode():
            emission, emission_lengths = self.align_model(wavs_tensors.to(self.device), wavs_lengths_tensor)
            star_dim = torch.zeros((emission.shape[0], emission.size(1), 1), dtype=emission.dtype, device=self.device)
            emission = torch.cat((emission, star_dim), dim=-1)
        
        logging.info("end alignment model forward")
        
        target_list = [self.get_target(transcript, language) for transcript, language in zip(transcript_list, language_list)]
        
        logging.info("align success")
        align_results = [
            self.align(emission_padded[:emission_length, :].unsqueeze(0), target)
            for emission_padded, emission_length, target in zip(emission, emission_lengths, target_list)
        ]
        
        logging.info("get align result")
        batch_aligned_tokens = [align_result[0] for align_result in align_results]
        batch_alignment_scores = [align_result[1] for align_result in align_results]

        alignment_result_list = [
            self.get_alignment_result(emission_padded, emission_length, aligned_tokens, alignment_scores, transcript, waveform, language)
            for emission_padded, emission_length, aligned_tokens, alignment_scores, transcript, waveform, language
            in zip(emission, emission_lengths, batch_aligned_tokens, batch_alignment_scores, transcript_list, wav_list, language_list)
        ]
        logging.info("get align result success")
        return alignment_result_list


async def batch_get_alignment_result_remote(alignment_url, audio_path, transcript, language):
    """
    通过调用远程对齐服务来批量获取对齐结果。
    """
    payload = {
        "audio_path": audio_path,
        "transcript": transcript,
        "language": language,
    }

    try:
        async with httpx.AsyncClient() as client:
            response = await client.post(alignment_url, json=payload, timeout=300) # 设置较长的超时
        response.raise_for_status()  # 如果状态码不是 2xx，则抛出异常
        data = response.json()
        return data['results']

    except requests.exceptions.RequestException as e:
        logging.error(f"Failed to connect to alignment service: {e}")
        traceback.print_exc()
        # 根据需求可以返回空列表或抛出异常
    except Exception as e:
        logging.error(f"An error occurred in remote alignment: {e}")
        traceback.print_exc()