jameson512
Upload files via LFS
1f235c3
Raw
History Blame Contribute Delete
12.5 kB
import json
import re
import time
from dataclasses import dataclass
from typing import List, Union
import requests
from videotrans.configure.excepts import StopRetry, SpeechToTextError
from videotrans.configure.config import tr, params, app_cfg, logger
from videotrans.recognition._base import BaseRecogn
from videotrans.task.taskcfg import SrtItem
from videotrans.util import tools
from videotrans.configure import contants
"""
请求发送:以二进制形式发送键名为 audio 的wav格式音频数据,采样率为16k、通道为1
requests.post(api_url, files={"audio": open(audio_file, 'rb')},data={"language":2位语言代码})
失败时返回
res={
"code":1,
"msg":"错误原因"
}
成功时返回
res={
"code":0,
"data":srt格式字符串
}
"""
RETRY_NUMS = 2
RETRY_DELAY = 10
@dataclass
class APIRecogn(BaseRecogn):
def __post_init__(self):
super().__post_init__()
api_url = params.get('recognapi_url', '').strip().rstrip('/').lower()
if not api_url.startswith('http'):
api_url = f'http://{api_url}'
if params.get('recognapi_key'):
if '?' in api_url:
api_url += f'&sk={params.get("recognapi_key", "")}'
else:
api_url += f'?sk={params.get("recognapi_key", "")}'
self.api_url = api_url
def _exec(self) -> Union[List[SrtItem], None]:
if self._exit(): return
if re.search(r'api\.gladia\.io', self.api_url, re.I):
return self._whisperzero()
if 'vibevoice-asr' in params.get('recognapi_key', ''):
return self._vibevoice_asr()
with open(self.audio_file, 'rb') as f:
chunk = f.read()
files = {"audio": chunk}
self.signal(
text=tr("Recognition may take a while, please be patient"))
res = requests.post(f"{self.api_url}", data={"language": self.detect_language}, files=files, timeout=1200)
res.raise_for_status()
content_type = res.headers.get('Content-Type','')
if 'application/json' not in content_type:
raise SpeechToTextError(res.text or res)
res = res.json()
if "code" not in res or res['code'] != 0:
raise SpeechToTextError(f'{res["msg"]}')
if "data" not in res or len(res['data']) < 1:
testdata={
"code":0,
"data":"SRT格式字符串"
}
testdata=json.dumps(testdata,ensure_ascii=False)
raise SpeechToTextError(f'识别出错,应返回类似数据:\n{testdata}\n\n但实际返回: {res}')
self.signal(
text=tools.get_srt_from_list(res['data']),
type='replace_subtitle'
)
if isinstance(res['data'],list):
data=[f'{i+1}\n{it["time"]}\n{it["text"]}' for i,it in enumerate(res['data'])]
data="\n\n".join(data)
else:
data=res['data']
return tools.get_subtitle_from_srt(data, is_file=False)
def _whisperzero(self)->Union[List[SrtItem], None]:
api_key = params.get("recognapi_key")
if not api_key:
raise SpeechToTextError(tr("api key must be filled in"))
# 上传 self.audio_file
with open(self.audio_file, "rb") as f:
audio_file = f.read()
files = {
"audio": (self.audio_file, audio_file, "audio/wav") # Content-Type 音频类型,有些API需要特别指定
}
response = requests.post("https://api.gladia.io/v2/upload", files=files, headers={
"x-gladia-key": api_key
})
response.raise_for_status()
audio_url = response.json()['audio_url']
payload = {
"detect_language": True if not self.detect_language or self.detect_language == 'auto' else False,
"enable_code_switching": False,
"language": "" if not self.detect_language or self.detect_language == 'auto' else self.detect_language[:2],
"subtitles": True,
"subtitles_config": {
"formats": ["srt"],
"minimum_duration": 1,
"maximum_duration": 15.5,
"maximum_characters_per_row": 80,
"maximum_rows_per_caption": 2,
"style": "default"
},
"sentences": True,
"punctuation_enhanced": True,
"audio_url": audio_url
}
response = requests.request("POST", 'https://api.gladia.io/v2/pre-recorded', json=payload, headers={
"x-gladia-key": api_key,
"Content-Type": "application/json"
})
response.raise_for_status()
id = response.json()['id']
# 获取结果
while 1:
if app_cfg.exit_soft: return
time.sleep(1)
response = requests.get(f"https://api.gladia.io/v2/pre-recorded/{id}", headers={"x-gladia-key": api_key})
response.raise_for_status()
d = response.json()
if d['status'] == 'error':
logger.warning(d)
raise StopRetry(f"Error:{d['error_code']}")
if d['status'] == 'done':
sens = d['result']['transcription']['subtitles'][0]['subtitles']
raws = tools.get_subtitle_from_srt(sens, is_file=False)
if self.detect_language and self.detect_language[:2] in contants.CJK_LANG:
for i, it in enumerate(raws):
text = re.sub(r'\s+', '', it['text'], flags=re.I | re.S)
raws[i]['text'] = text
return raws
def _vibevoice_asr(self)->Union[List[SrtItem], None]:
from gradio_client import Client, handle_file
from pydub import AudioSegment
import re
import ast
import os
import json
from pathlib import Path
# 定义切片时长 (60分钟 = 60 * 60 * 1000 毫秒)
CHUNK_DURATION_MS = 60 * 60 * 1000
# 初始化客户端
client = Client(self.api_url, httpx_kwargs={"timeout": 7200})
# 内部函数:处理单个片段的返回结果
def _process_chunk_result(raw_text, time_offset_ms, start_line_index):
# 1. 使用正则表达式找到列表部分
match = re.search(r'(\[{.*?}])', raw_text, re.DOTALL)
chunk_raws = []
chunk_speaker_raw_list = [] # 仅收集当前片段的原始说话人标记
if not match:
# 如果某个片段没识别出内容(可能是静音),返回空而不是报错
logger.warning(f"No subtitles found in chunk starting at {time_offset_ms}ms")
return [], []
list_str = match.group(1)
list_str = re.sub(r'^.*?\[{', '[{', list_str, flags=re.S)
list_str = re.sub(r'}].*$', '}]', list_str, flags=re.S)
list_str = re.sub(r"\n?\n", '', list_str)
segments = None
try:
segments = json.loads(list_str)
except json.JSONDecodeError:
try:
segments = ast.literal_eval(list_str)
except (ValueError, SyntaxError):
context = {
"null": None,
"true": True,
"false": False,
"__builtins__": None
}
segments = eval(list_str, context)
except Exception as e:
logger.error(f"AST eval failed: {e}")
if not segments:
return [], []
# 2. 遍历结果并加上时间偏移
for i, seg in enumerate(segments):
# 计算加上偏移量后的毫秒数
seg_start_ms = int(float(seg['Start']) * 1000) + time_offset_ms
seg_end_ms = int(float(seg['End']) * 1000) + time_offset_ms
tmp = {
"line": start_line_index + i + 1, # 累加行号
"text": seg['Content'],
"start_time": seg_start_ms,
"end_time": seg_end_ms,
}
# [Noise]之类无有效信息
if re.match(r'^\[[a-zA-Z0-9\s]+]$', seg['Content'].strip()):
continue
# 假设 tools 是你类外部或全局可访问的工具
tmp['startraw'] = tools.ms_to_time_string(ms=tmp['start_time'])
tmp['endraw'] = tools.ms_to_time_string(ms=tmp['end_time'])
tmp['time'] = f"{tmp['startraw']} --> {tmp['endraw']}"
chunk_raws.append(tmp)
# 收集原始说话人信息 (例如 "Speaker 1")
sp = seg.get("Speaker", '-')
chunk_speaker_raw_list.append(sp)
return chunk_raws, chunk_speaker_raw_list
# self.audio_file 是 wav 路径
audio = AudioSegment.from_wav(self.audio_file)
total_duration = len(audio)
final_raws = []
all_speaker_raw_list = [] # 存储所有片段原本的说话人标记
current_line = 0
for i, start_ms in enumerate(range(0, total_duration, CHUNK_DURATION_MS)):
end_ms = min(start_ms + CHUNK_DURATION_MS, total_duration)
# 切割音频
chunk_audio = audio[start_ms:end_ms]
# 保存临时文件
temp_chunk_path = os.path.join(self.cache_folder, f"temp_chunk_{i}.wav")
chunk_audio.export(temp_chunk_path, format="wav")
try:
result = client.predict(
audio_input=handle_file(temp_chunk_path),
audio_path_input=None,
start_time_input=None,
end_time_input=None,
max_new_tokens=65536,
temperature=0,
top_p=1,
do_sample=False,
repetition_penalty=1,
context_info="",
api_name="/transcribe_audio"
)
# 处理返回结果,传入当前的 start_ms 作为时间偏移量
chunk_data, chunk_spk = _process_chunk_result(
result[0],
time_offset_ms=start_ms,
start_line_index=current_line
)
final_raws.extend(chunk_data)
all_speaker_raw_list.extend(chunk_spk)
current_line += len(chunk_data)
except Exception as e:
logger.exception(f"Error processing chunk {i}: {e}")
finally:
# 清理临时文件
if os.path.exists(temp_chunk_path):
os.remove(temp_chunk_path)
if not final_raws:
raise SpeechToTextError(f'VibeVoice:{self.api_url} not return data')
# 统一处理说话人逻辑 (合并后的重排序)
# 这里是将所有片段的说话人混在一起处理。
# 警告:VibeVoice 是分段处理的,Chunk1 的 spk0 和 Chunk2 的 spk0 可能不是同一个人。
final_speaker_list = []
unique_speakers = []
# 提取不重复的说话人列表保持顺序
for sp in all_speaker_raw_list:
if sp not in unique_speakers:
unique_speakers.append(sp)
if unique_speakers:
try:
# 生成最终的 spk0, spk1... 映射
for sp in all_speaker_raw_list:
if sp == '-':
# 如果没有识别出,暂定为最后一个新编号
final_speaker_list.append(f'spk{len(unique_speakers)}')
else:
final_speaker_list.append(f'spk{unique_speakers.index(sp)}')
# 写入最终的 speaker.json
if final_speaker_list:
Path(f'{self.cache_folder}/speaker.json').write_text(json.dumps(final_speaker_list),
encoding='utf-8')
except Exception as e:
logger.exception(f'说话人重排序出错,忽略{e}', exc_info=True)
return final_raws