add: full example of forced alignment for music

.gitignore

@@ -16,4 +16,5 @@ translate/result
 dataset/raw
 translate/special-spiders
 ugNMT/BPE/output*
-ugNMT/BPE/codes
+ugNMT/BPE/codes
+forced-alignment/segments

forced-alignment/1.fff

@@ -0,0 +1,82 @@
+[0]: 25.462-27.102
+[1]: 27.202-28.382
+[2]: 28.582-30.262
+[3]: 30.622-32.303
+[4]: 32.403-33.643
+[5]: 33.703-35.483
+[6]: 35.843-37.503
+[7]: 37.643-38.803
+[8]: 38.903-40.563
+[9]: 41.023-42.683
+[10]: 42.803-43.883
+[11]: 44.063-45.884
+[12]: 56.584-58.245
+[13]: 58.345-59.585
+[14]: 59.685-61.345
+[15]: 61.785-63.405
+[16]: 63.525-64.785
+[17]: 64.825-66.565
+[18]: 66.925-68.665
+[19]: 68.785-69.826
+[20]: 70.005-71.686
+[21]: 72.106-73.846
+[22]: 73.946-75.086
+[23]: 75.226-76.986
+[24]: 77.346-79.526
+[25]: 79.646-81.426
+[26]: 81.566-83.527
+[27]: 84.307-86.847
+[28]: 86.927-89.487
+[29]: 89.527-91.947
+[30]: 92.107-94.667
+[31]: 94.707-97.128
+[32]: 97.288-99.848
+[33]: 99.908-102.328
+[34]: 102.448-116.229
+[35]: 116.249-117.949
+[36]: 118.029-119.129
+[37]: 119.369-121.07
+[38]: 121.45-123.13
+[39]: 123.23-124.47
+[40]: 124.53-126.27
+[41]: 126.67-128.29
+[42]: 128.41-129.59
+[43]: 129.71-131.51
+[44]: 131.81-133.511
+[45]: 133.611-134.671
+[46]: 134.891-136.651
+[47]: 137.011-138.671
+[48]: 138.791-139.991
+[49]: 140.091-141.811
+[50]: 142.191-143.871
+[51]: 143.971-145.071
+[52]: 145.271-146.952
+[53]: 147.372-149.052
+[54]: 149.132-150.412
+[55]: 150.492-152.212
+[56]: 152.552-154.272
+[57]: 154.372-155.492
+[58]: 155.652-157.412
+[59]: 157.792-159.953
+[60]: 160.093-161.873
+[61]: 162.013-163.953
+[62]: 164.733-167.293
+[63]: 167.353-169.913
+[64]: 169.953-172.394
+[65]: 172.534-175.114
+[66]: 175.154-177.594
+[67]: 177.714-180.294
+[68]: 180.314-182.854
+[69]: 182.934-185.575
+[70]: 195.875-198.436
+[71]: 198.476-201.056
+[72]: 201.096-203.516
+[73]: 203.656-206.196
+[74]: 206.276-208.796
+[75]: 208.857-211.417
+[76]: 211.457-213.917
+[77]: 214.057-216.617
+[78]: 216.677-219.077
+[79]: 219.217-221.817
+[80]: 221.817-224.378
+[81]: 224.458-227.098

forced-alignment/1.lrc

@@ -0,0 +1,86 @@
+[00:25.513] 东汉末狼烟不休
+[00:27.178] 常侍乱 朝野陷
+[00:28.499] 阿瞒挟天子 令诸侯
+[00:30.838] 踞江东志在九州
+[00:32.457] 继祖业 承父兄
+[00:33.793] 既冕主吴越 万兜鍪
+[00:35.878] 纵天下几变春秋
+[00:37.536] 稳东南 面中原
+[00:38.839] 水师锁长江 抗曹刘
+[00:40.978] 镇赤壁雄风赳赳
+[00:42.872] 夺荆楚 抚山越
+[00:44.043] 驱金戈铁马 灭仇雠
+[00:46.740]
+[00:56.690] 紫发髯碧色眼眸
+[00:58.324] 射猛虎 倚黄龙
+[00:59.694] 胆识过凡人 谁敌手
+[01:01.723] 御天下半百之久
+[01:03.624] 选贤臣 任能将
+[01:04.865] 覆江东云雨 尽风流
+[01:07.034] 千秋过再难回首
+[01:08.819] 问古今 兴亡事
+[01:10.059] 几人耀青史 芳名留
+[01:12.075] 笑谈间云烟已旧
+[01:13.915] 终留下 万古叹
+[01:15.278] 生子该当如 孙仲谋
+[01:17.458] 运帷幄 英雄几拂袖
+[01:19.644] 阴谋 阳谋 明仇 暗斗
+[01:21.637] 化作一江浊浪东流
+[01:24.443] 君不见军赤壁纵野火铁索连环
+[01:26.739] 也不见御北敌联西蜀长江上鏖战
+[01:29.596] 继遗志领江东屹立于神州东南
+[01:32.188] 尽心力洒英血展伟业剑气指苍天
+[01:34.840] 军帐内公瑾智张昭谋奇策频献
+[01:37.241] 沙场上太史勇甘宁霸一骑当十千
+[01:39.965] 纵使有千万种寂寞和孤单相伴
+[01:42.425] 既受终冠帝冕龙椅上成败也笑看
+[01:46.351]
+[01:56.184] 铁瓮城难攻易守
+[01:58.069] 旌旗立 苍空蔽
+[01:59.438] 逾百千雄师 万蒙舟
+[02:01.462] 善制衡眼光独秀
+[02:03.301] 擢鲁肃 劝阿蒙
+[02:04.557] 聚贤成霸业 名利收
+[02:06.500] 固疆土施德恩厚
+[02:08.302] 军心定 百姓安
+[02:09.712] 富国又强兵 重耕耨
+[02:11.783] 交远好未雨绸缪
+[02:13.642] 联南洋 合林邑
+[02:14.914] 行军远渡海 驻夷洲
+[02:16.944] 残垣下枯木雕朽
+[02:18.784] 想当年 麦城边
+[02:20.076] 截兵缚关羽 终其寿
+[02:22.263] 凭栏倚横看吴钩
+[02:23.993] 叹乱世 几时了
+[02:25.398] 天下归一统 没其咎
+[02:27.462] 称帝王壮心仍稠
+[02:29.061] 却无奈 自孤傲
+[02:30.438] 同室亦操戈 子嗣斗
+[02:32.512] 千年后恚恨徒留
+[02:34.291] 再何寻 军帐里
+[02:35.698] 将士聚欢饮 赏箜篌
+[02:37.763] 运帷幄 英雄几拂袖
+[02:40.029] 阴谋 阳谋 明仇 暗斗
+[02:42.146] 化作一江浊浪东流
+[02:44.641] 君不见吕子明踏轻舟白衣渡川
+[02:47.318] 也不见陆伯言烧联营火光上冲天
+[02:49.870] 善制衡选贤臣任能将共谋江山
+[02:52.410] 听忠言摒逆语树威严宝剑斫书案
+[02:55.117] 夺荆州抗刘备合曹操共克襄樊
+[02:57.692] 守夷陵任陆逊剿敌军 火计破蜀胆
+[03:00.391] 固江河成帝业立国家终归于乱
+[03:02.864] 光阴逝千载过功成者都付笑谈间
+[03:06.250]
+[03:15.700] 君不见军赤壁纵野火铁索连环
+[03:18.545] 也不见御北敌联西蜀长江上鏖战
+[03:21.080] 继遗志领江东屹立于神州东南
+[03:23.649] 尽心力洒英血展伟业剑指苍天
+[03:26.263] 君不见吕子明踏轻舟白衣渡川
+[03:28.746] 也不见陆伯言烧联营火光上冲天
+[03:31.395] 善制衡选贤臣任能将共谋江山
+[03:34.021] 听忠言摒逆语树威严宝剑斫书案
+[03:36.589] 纵使有千万种寂寞和孤单相伴
+[03:39.103] 既受终冠帝冕龙椅上成败也笑看
+[03:41.916] 固江河成帝业立国家终归于乱
+[03:44.470] 光阴逝千载过功成者都付笑谈间
+[03:48.436]

forced-alignment/1.txt

@@ -0,0 +1,82 @@
+东汉末狼烟不休
+常侍乱朝野陷
+阿瞒挟天子令诸侯
+踞江东志在九州
+继祖业承父兄
+既冕主吴越万兜鍪
+纵天下几变春秋
+稳东南面中原
+水师锁长江抗曹刘
+镇赤壁雄风赳赳
+夺荆楚抚山越
+驱金戈铁马灭仇雠
+紫发髯碧色眼眸
+射猛虎倚黄龙
+胆识过凡人谁敌手
+御天下半百之久
+选贤臣任能将
+覆江东云雨尽风流
+千秋过再难回首
+问古今兴亡事
+几人耀青史芳名留
+笑谈间云烟已旧
+终留下万古叹
+生子该当如孙仲谋
+运帷幄英雄几拂袖
+阴谋阳谋明仇暗斗
+化作一江浊浪东流
+君不见军赤壁纵野火铁索连环
+也不见御北敌联西蜀长江上鏖战
+继遗志领江东屹立于神州东南
+尽心力洒英血展伟业剑气指苍天
+军帐内公瑾智张昭谋奇策频献
+沙场上太史勇甘宁霸一骑当十千
+纵使有千万种寂寞和孤单相伴
+既受终冠帝冕龙椅上成败也笑看
+铁瓮城难攻易守
+旌旗立苍空蔽
+逾百千雄师万蒙舟
+善制衡眼光独秀
+擢鲁肃劝阿蒙
+聚贤成霸业名利收
+固疆土施德恩厚
+军心定百姓安
+富国又强兵重耕耨
+交远好未雨绸缪
+联南洋合林邑
+行军远渡海驻夷洲
+残垣下枯木雕朽
+想当年麦城边
+截兵缚关羽终其寿
+凭栏倚横看吴钩
+叹乱世几时了
+天下归一统没其咎
+称帝王壮心仍稠
+却无奈自孤傲
+同室亦操戈子嗣斗
+千年后恚恨徒留
+再何寻军帐里
+将士聚欢饮赏箜篌
+运帷幄英雄几拂袖
+阴谋阳谋明仇暗斗
+化作一江浊浪东流
+君不见吕子明踏轻舟白衣渡川
+也不见陆伯言烧联营火光上冲天
+善制衡选贤臣任能将共谋江山
+听忠言摒逆语树威严宝剑斫书案
+夺荆州抗刘备合曹操共克襄樊
+守夷陵任陆逊剿敌军火计破蜀胆
+固江河成帝业立国家终归于乱
+光阴逝千载过功成者都付笑谈间
+君不见军赤壁纵野火铁索连环
+也不见御北敌联西蜀长江上鏖战
+继遗志领江东屹立于神州东南
+尽心力洒英血展伟业剑指苍天
+君不见吕子明踏轻舟白衣渡川
+也不见陆伯言烧联营火光上冲天
+善制衡选贤臣任能将共谋江山
+听忠言摒逆语树威严宝剑斫书案
+纵使有千万种寂寞和孤单相伴
+既受终冠帝冕龙椅上成败也笑看
+固江河成帝业立国家终归于乱
+光阴逝千载过功成者都付笑谈间

forced-alignment/split.py

@@ -0,0 +1,179 @@
+import torch
+import torchaudio
+from typing import List
+from pypinyin import lazy_pinyin
+from pypinyin_dict.phrase_pinyin_data import cc_cedict
+from torchaudio.transforms import Resample
+from tqdm import tqdm 
+
+
+def compute_alignments(waveform: torch.Tensor, transcript: List[str]):
+    with torch.inference_mode():
+        emission, _ = model(waveform.to(device))
+        token_spans = aligner(emission[0], tokenizer(transcript))
+    return emission, token_spans
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+from torchaudio.pipelines import MMS_FA as bundle
+
+model = bundle.get_model()
+model.to(device)
+
+tokenizer = bundle.get_tokenizer()
+aligner = bundle.get_aligner()
+
+cc_cedict.load()
+
+add_spaces = lambda s: ' '.join(s)
+
+from pydub import AudioSegment
+
+def get_audio_duration(file_path):
+    """
+    读取音频文件并获取其时长（秒数）。
+
+    :param file_path: 音频文件的路径
+    :return: 音频文件的时长（秒数）
+    """
+    try:
+        audio = AudioSegment.from_file(file_path)
+        duration_in_seconds = len(audio) / 1000.0
+        return duration_in_seconds
+    except Exception as e:
+        print(f"Error reading audio file: {e}")
+        return None
+
+def timestamp(seconds):
+    """
+    将浮点数的秒钟转换为TTML的时间戳格式（HH:MM:SS.sss）。
+
+    :param seconds: 浮点数的秒钟
+    :return: TTML时间戳格式字符串
+    """
+    hours = int(seconds // 3600)
+    minutes = int((seconds % 3600) // 60)
+    seconds = seconds % 60
+    milliseconds = int((seconds % 1) * 1000)
+    seconds = int(seconds)
+
+    return f"{hours:02}:{minutes:02}:{seconds:02}.{milliseconds:03}"
+
+import xml.etree.ElementTree as ET
+
+import os
+import re
+
+def extract_numbers_from_files(directory):
+    """
+    读取给定目录，提取文件名中的数字部分，并返回一个包含这些数字的列表。
+
+    :param directory: 目录路径
+    :return: 包含数字的列表
+    """
+    numbers = []
+    pattern = re.compile(r'line-(\d+)\.wav')
+
+    try:
+        for filename in os.listdir(directory):
+            match = pattern.match(filename)
+            if match:
+                number = int(match.group(1))
+                numbers.append(number)
+    except Exception as e:
+        print(f"Error reading directory: {e}")
+        return None
+
+    return numbers
+    
+class TTMLGenerator:
+    def __init__(self, duration, xmlns="http://www.w3.org/ns/ttml", xmlns_ttm="http://www.w3.org/ns/ttml#metadata", xmlns_amll="http://www.example.com/ns/amll", xmlns_itunes="http://music.apple.com/lyric-ttml-internal"):
+        self.tt = ET.Element("tt", attrib={
+            "xmlns": xmlns,
+            "xmlns:ttm": xmlns_ttm,
+            "xmlns:amll": xmlns_amll,
+            "xmlns:itunes": xmlns_itunes
+        })
+        self.head = ET.SubElement(self.tt, "head")
+        self.metadata = ET.SubElement(self.head, "metadata")
+        self.body = ET.SubElement(self.tt, "body", attrib={"dur": duration})
+        self.div = ET.SubElement(self.body, "div")
+
+    def add_lyrics(self, begin, end, agent, itunes_key, words):
+        p = ET.SubElement(self.div, "p", attrib={
+            "begin": begin,
+            "end": end,
+            "ttm:agent": agent,
+            "itunes:key": itunes_key
+        })
+        for word, start, stop in words:
+            span = ET.SubElement(p, "span", attrib={"begin": start, "end": stop})
+            span.text = word
+
+    def save(self, filename):
+        tree = ET.ElementTree(self.tt)
+        tree.write(filename, encoding="utf-8", xml_declaration=True)
+
+duration = get_audio_duration("霜雪千年_vocal.mp3")
+
+# 示例使用
+ttml_generator = TTMLGenerator(duration=timestamp(duration))
+
+
+def process_line(line_idx, start_time):
+    with open(f"./segments/line-{line_idx}.txt", "r") as f:
+        text = f.read()
+        
+    waveform, sample_rate = torchaudio.load(f"./segments/line-{line_idx}.wav")
+
+    waveform = waveform[0:1]
+    resampler = Resample(orig_freq=sample_rate, new_freq=16000)
+    waveform = resampler(waveform)
+
+    text_pinyin = lazy_pinyin(text)
+    text_normalized = " ".join(text_pinyin)
+    
+    transcript = text_normalized.split()
+    emission, token_spans = compute_alignments(waveform, transcript)
+    num_frames = emission.size(1)
+    ratio = waveform.size(1) / num_frames
+
+    words = []
+    for i in range(len(token_spans)):
+        spans = token_spans[i]
+        x0 = start_time + int(ratio * spans[0].start) / 16000
+        x1 = start_time + int(ratio * spans[-1].end) / 16000
+        words.append({
+            "word": text[i],
+            "start": x0,
+            "end": x1
+        })
+    idx=0
+    for item in words:
+        if idx == len(words) - 1:
+            break
+        item["end"] = words[idx + 1]["start"]
+        idx+=1
+    result = []
+    for word in words:
+        result.append((word["word"], timestamp(word["start"]), timestamp(word["end"])))
+    return result
+
+
+lines_to_process = sorted(extract_numbers_from_files("segments"))
+
+i=1
+for line_num in tqdm(lines_to_process):
+    with open(f"./segments/line-{line_num}.time", "r") as f:
+        a = f.read()
+        b = a.split(",")
+        start_time = float(b[0])
+        end_time = float(b[1])
+    result = process_line(line_num, start_time)
+    ttml_generator.add_lyrics(
+        begin=timestamp(start_time), end=timestamp(end_time), agent="v1", itunes_key=f"L{i}",
+        words=result
+    )
+    i+=1
+
+# 保存文件
+ttml_generator.save("output.ttml")

forced-alignment/split_song.py

@@ -0,0 +1,57 @@
+from pydub import AudioSegment
+import re
+
+def parse_lrc(lrc_file):
+    """解析LRC文件，返回一个包含时间戳和歌词的列表"""
+    with open(lrc_file, 'r', encoding='utf-8') as f:
+        lines = f.readlines()
+    
+    lrc_data = []
+    for line in lines:
+        # 使用正则表达式匹配时间戳和歌词
+        match = re.match(r'\[(\d+):(\d+\.\d+)\](.*)', line)
+        if match:
+            minutes = int(match.group(1))
+            seconds = float(match.group(2))
+            lyric = match.group(3).strip()
+            lyric = lyric.replace(" ", "")
+            timestamp = minutes * 60 + seconds
+            lrc_data.append((timestamp, lyric))
+    
+    return lrc_data
+
+def split_audio_by_lrc(audio_file, lrc_data, output_prefix):
+    """根据LRC数据分割音频文件，并保存为单独的WAV文件"""
+    audio = AudioSegment.from_file(audio_file)
+    
+    for i, (start_time, lyric) in enumerate(lrc_data):
+        # Skip empty line
+        if lyric.strip() == "":
+            continue
+        if i < len(lrc_data) - 1:
+            end_time = lrc_data[i + 1][0]
+        else:
+            end_time = len(audio) / 1000  # 最后一行歌词到音频结束
+        start_time = max(0, start_time - 0.1)  # 前后各扩0.1秒
+        end_time = min(len(audio) / 1000, end_time + 0.1)
+        start_time_ms = start_time * 1000
+        end_time_ms = end_time * 1000
+        
+        segment = audio[start_time_ms:end_time_ms]
+        output_file = f"{output_prefix}-{i+1}.wav"
+        output_script = f"{output_prefix}-{i+1}.txt"
+        output_time = f"{output_prefix}-{i+1}.time"
+        segment.export(output_file, format="wav")
+        with open(output_script, "w") as f:
+            f.write(lyric)
+        with open(output_time, "w") as f:
+            f.write(str(start_time)+","+str(end_time))
+        print(f"Saved {output_file}")
+
+if __name__ == "__main__":
+    lrc_file = "霜雪千年.lrc"  # LRC文件路径
+    audio_file = "霜雪千年.mp3"  # 音频文件路径
+    output_prefix = "segments/line"  # 输出文件名的前缀
+    
+    lrc_data = parse_lrc(lrc_file)
+    split_audio_by_lrc(audio_file, lrc_data, output_prefix)

forced-alignment/split_whole.py

@@ -0,0 +1,122 @@
+import torch
+import torchaudio
+from typing import List
+from pypinyin import lazy_pinyin
+from pypinyin_dict.phrase_pinyin_data import cc_cedict
+from torchaudio.transforms import Resample
+import xml.etree.ElementTree as ET
+
+from pydub import AudioSegment
+
+def get_audio_duration(file_path):
+    """
+    读取音频文件并获取其时长（秒数）。
+
+    :param file_path: 音频文件的路径
+    :return: 音频文件的时长（秒数）
+    """
+    try:
+        audio = AudioSegment.from_file(file_path)
+        duration_in_seconds = len(audio) / 1000.0
+        return duration_in_seconds
+    except Exception as e:
+        print(f"Error reading audio file: {e}")
+        return None
+
+def timestamp(seconds):
+    """
+    将浮点数的秒钟转换为TTML的时间戳格式（HH:MM:SS.sss）。
+
+    :param seconds: 浮点数的秒钟
+    :return: TTML时间戳格式字符串
+    """
+    hours = int(seconds // 3600)
+    minutes = int((seconds % 3600) // 60)
+    seconds = seconds % 60
+    milliseconds = int((seconds % 1) * 1000)
+    seconds = int(seconds)
+
+    return f"{hours:02}:{minutes:02}:{seconds:02}.{milliseconds:03}"
+
+class TTMLGenerator:
+    def __init__(self, duration, xmlns="http://www.w3.org/ns/ttml", xmlns_ttm="http://www.w3.org/ns/ttml#metadata", xmlns_amll="http://www.example.com/ns/amll", xmlns_itunes="http://music.apple.com/lyric-ttml-internal"):
+        self.tt = ET.Element("tt", attrib={
+            "xmlns": xmlns,
+            "xmlns:ttm": xmlns_ttm,
+            "xmlns:amll": xmlns_amll,
+            "xmlns:itunes": xmlns_itunes
+        })
+        self.head = ET.SubElement(self.tt, "head")
+        self.metadata = ET.SubElement(self.head, "metadata")
+        self.body = ET.SubElement(self.tt, "body", attrib={"dur": duration})
+        self.div = ET.SubElement(self.body, "div")
+
+    def add_lyrics(self, begin, end, agent, itunes_key, words):
+        p = ET.SubElement(self.div, "p", attrib={
+            "begin": begin,
+            "end": end,
+            "ttm:agent": agent,
+            "itunes:key": itunes_key
+        })
+        for word, start, stop in words:
+            span = ET.SubElement(p, "span", attrib={"begin": start, "end": stop})
+            span.text = word
+
+    def save(self, filename):
+        tree = ET.ElementTree(self.tt)
+        tree.write(filename, encoding="utf-8", xml_declaration=True)
+
+
+def compute_alignments(waveform: torch.Tensor, transcript: List[str]):
+    with torch.inference_mode():
+        emission, _ = model(waveform.to(device))
+        token_spans = aligner(emission[0], tokenizer(transcript))
+    return emission, token_spans
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+from torchaudio.pipelines import MMS_FA as bundle
+
+model = bundle.get_model()
+model.to(device)
+
+tokenizer = bundle.get_tokenizer()
+aligner = bundle.get_aligner()
+
+cc_cedict.load()
+
+add_spaces = lambda s: ' '.join(s)
+
+with open("./1.txt", "r") as f:
+    text_lines = f.readlines()
+
+text_pinyin = []
+
+for line in text_lines:
+    text_pinyin.append("".join(lazy_pinyin(line.strip())))
+
+text_normalized = " ".join(text_pinyin)
+
+print(text_normalized)
+
+waveform, sample_rate = torchaudio.load("./权御天下 [vocals].mp3")
+
+waveform = waveform[0:1]
+resampler = Resample(orig_freq=sample_rate, new_freq=16000)
+waveform = resampler(waveform)
+
+transcript = text_normalized.split()
+emission, token_spans = compute_alignments(waveform, transcript)
+num_frames = emission.size(1)
+
+ratio = waveform.size(1) / num_frames
+
+duration = get_audio_duration("权御天下 [vocals].mp3")
+
+ttml_generator = TTMLGenerator(duration=timestamp(duration))
+
+for i in range(len(token_spans)):
+    spans = token_spans[i]
+    x0 = round(int(ratio * spans[0].start) / 16000, 3)
+    x1 = round(int(ratio * spans[-1].end) / 16000, 3)
+    with open("1.fff", "a") as f:
+        f.write(f"{[i]}: {x0}-{x1}\n")

forced-alignment/test_split.py

@@ -0,0 +1,60 @@
+import torch
+import torchaudio
+from typing import List
+from pypinyin import lazy_pinyin
+from pypinyin_dict.phrase_pinyin_data import cc_cedict
+from torchaudio.transforms import Resample
+
+
+def compute_alignments(waveform: torch.Tensor, transcript: List[str]):
+    with torch.inference_mode():
+        emission, _ = model(waveform.to(device))
+        token_spans = aligner(emission[0], tokenizer(transcript))
+    return emission, token_spans
+
+# Compute average score weighted by the span length
+def _score(spans):
+    return sum(s.score * len(s) for s in spans) / sum(len(s) for s in spans)
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+from torchaudio.pipelines import MMS_FA as bundle
+
+model = bundle.get_model()
+model.to(device)
+
+tokenizer = bundle.get_tokenizer()
+aligner = bundle.get_aligner()
+
+cc_cedict.load()
+
+add_spaces = lambda s: ' '.join(s)
+
+with open("./segments/line-1.txt", "r") as f:
+    text = f.read()
+
+text_raw = add_spaces(text)
+text_list = list(text)
+text_pinyin = lazy_pinyin(text)
+text_normalized = " ".join(text_pinyin)
+
+waveform, sample_rate = torchaudio.load("./segments/line-1.wav")
+
+waveform = waveform[0:1]
+resampler = Resample(orig_freq=sample_rate, new_freq=16000)
+waveform = resampler(waveform)
+
+transcript = text_normalized.split()
+emission, token_spans = compute_alignments(waveform, transcript)
+num_frames = emission.size(1)
+
+
+print("Raw Transcript: ", text_raw)
+print("Normalized Transcript: ", text_normalized)
+
+ratio = waveform.size(1) / num_frames
+
+for i in range(len(token_spans)):
+    spans = token_spans[i]
+    x0 = round(int(ratio * spans[0].start) / 16000, 3)
+    x1 = round(int(ratio * spans[-1].end) / 16000, 3)
+    print(f"{text[i]}: {x0}-{x1}")

forced-alignment/霜雪千年.lrc

@@ -0,0 +1,55 @@
+[tool: 歌词滚动姬 https://lrc-maker.github.io]
+[ti: 霜雪千年]
+[ar: COP]
+[00:13.876] 梨花香缠着衣角掠过熙攘
+[00:19.385] 复悄入红帘深帐
+[00:22.775] 听枝头黄鹂逗趣儿细风绕指淌
+[00:27.437] 坐船舫兰桨拨开雾霭迷茫
+[00:33.000] 不觉已一日过半
+[00:36.361] 过眼的葱郁风光悉数泛了黄
+[00:41.590] 褪尽温度的风无言牵引中
+[00:44.898] 便清晰了在此的眉目
+[00:48.721] 暮色的消融隐约了晦朔葱茏
+[00:53.888] 在这老街回眸烟云中追溯我是谁
+[00:58.582] 只消暮雨点滴便足以粉饰这是非
+[01:02.158] 待这月色涌起谁人轻叩这门扉
+[01:09.000] 苔绿青石板街
+[01:10.444] 斑驳了流水般岁月
+[01:12.389] 小酌三盏两杯理不清缠绕的情结
+[01:15.868] 在你淡漠眉间
+[01:19.277] 瞥见离人的喜悲霜雪
+[01:25.766]
+[01:36.176] 楼阁现尘飞雾散荧光蹁跹
+[01:41.500] 显露出斑驳石阶
+[01:44.865] 入眼是落英纷然芳草入深院
+[01:49.571] 凭栏杆小桌上置琼觞两盏
+[01:55.277] 阖眼听清风疏叶
+[01:58.603] 似曾有欢声笑言萦绕这高轩
+[02:03.691] 云动寂静鸣蝉雨坠激漪涟
+[02:07.298] 皴擦点染勾勒这世间
+[02:11.084] 缘起的一眼定格了三生千年
+[02:16.245] 在这老街回眸烟云中追溯我是谁
+[02:20.988] 只消暮雨点滴便足以粉饰这是非
+[02:24.463] 待这月色涌起谁人轻叩这门扉
+[02:31.390] 苔绿青石板街
+[02:32.752] 斑驳了流水般岁月
+[02:34.687] 小酌三盏两杯理不清缠绕的情结
+[02:38.162] 在你淡漠眉间
+[02:41.495] 瞥见离人的喜悲霜雪
+[02:48.066]
+[02:58.863] 三月梨花雪几载开了又败
+[03:02.206] 笔锋走黑白丹青中穿插无奈
+[03:05.594] 彼时那弯月何时初现于江畔
+[03:08.966] 而我又在待何人
+[03:11.201] 在这亭台回眸千年后
+[03:14.508] 忆起你是谁
+[03:15.816] 只消月色隐约便足以勾勒这是非
+[03:19.323] 待这回忆涌起恍惚之间已下泪
+[03:26.171] 枫红十里长街
+[03:27.662] 红帘后谁人蹙着眉
+[03:29.608] 遥梦桑竹桃源
+[03:31.188] 轮回中曾道别的地点
+[03:32.832] 愿今生再相见
+[03:36.478] 消融你眉间
+[03:38.585] 悲戚霜雪
+[03:44.184]