Ma/auto-caption
1
0
Fork 0
mirror of https://github.com/HiMeditator/auto-caption.git synced 2026-02-04 04:14:42 +08:00
Code Issues Packages Projects Releases Wiki Activity

feat(engine): 重构字幕引擎，新增 Sherpa-ONNX SenseVoice 语音识别模型

Browse Source 
- 重构字幕引擎，将音频采集改为在新线程上进行
- 重构 audio2text 中的类，调整运行逻辑
- 更新 main 函数，添加对 Sosv 模型的支持
- 修改 AudioStream 类，默认使用 16000Hz 采样率
This commit is contained in:
himeditator
2025-09-06 20:49:46 +08:00
parent 2b7ce06f04
commit eba2c5ca45
14 changed files with 377 additions and 112 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
engine/audio2text/__init__.py
View File

@@ -1,3 +1,3 @@
from dashscope.common.error import InvalidParameter
from .gummy import GummyRecognizer
from .vosk import VoskRecognizer
from .sosv import SosvRecognizer

25
engine/audio2text/gummy.py
View File

@@ -5,9 +5,10 @@ from dashscope.audio.asr import (
TranslationRecognizerRealtime
)
import dashscope
from dashscope.common.error import InvalidParameter
from datetime import datetime
from utils import stdout_cmd, stdout_obj, stderr
from utils import stdout_cmd, stdout_obj, stdout_err
from utils import shared_data
class Callback(TranslationRecognizerCallback):
"""
@@ -90,9 +91,23 @@ class GummyRecognizer:
"""启动 Gummy 引擎"""
self.translator.start()
def send_audio_frame(self, data):
"""发送音频帧，擎将自动识别并将识别结果输出到标准输出中"""
self.translator.send_audio_frame(data)
def translate(self):
"""持续读取共享数据中的音频帧，并进行语音识别，将识别结果输出到标准输出中"""
global shared_data
restart_count = 0
while shared_data.status == 'running':
chunk = shared_data.chunk_queue.get()
try:
self.translator.send_audio_frame(chunk)
except InvalidParameter as e:
restart_count += 1
if restart_count > 5:
stdout_err(str(e))
shared_data.status = "kill"
stdout_cmd('kill')
break
else:
stdout_cmd('info', f'Gummy engine stopped, restart attempt: {restart_count}...')
def stop(self):
"""停止 Gummy 引擎"""

139
engine/audio2text/sosv.py Normal file
View File

@@ -0,0 +1,139 @@
"""
Shepra-ONNX SenseVoice Model
This code file references the following:
https://github.com/k2-fsa/sherpa-onnx/blob/master/python-api-examples/simulate-streaming-sense-voice-microphone.py
"""
import time
from datetime import datetime
import sherpa_onnx
import numpy as np
from utils import shared_data
from utils import stdout_cmd, stdout_obj
from utils import google_translate, ollama_translate
class SosvRecognizer:
"""
使用 Sense Voice 非流式模型处理流式音频数据，并在标准输出中输出 Auto Caption 软件可读取的 JSON 字符串数据
初始化参数：
model_path: Shepra ONNX Sense Voice 识别模型路径
vad_model: Silero VAD 模型路径
target: 翻译目标语言
trans_model: 翻译模型名称
ollama_name: Ollama 模型名称
"""
def __init__(self, model_path: str, target: str | None, trans_model: str, ollama_name: str):
if model_path.startswith('"'):
model_path = model_path[1:]
if model_path.endswith('"'):
model_path = model_path[:-1]
self.model_path = model_path
self.target = target
if trans_model == 'google':
self.trans_func = google_translate
else:
self.trans_func = ollama_translate
self.ollama_name = ollama_name
self.time_str = ''
self.cur_id = 0
self.prev_content = ''
def start(self):
"""启动 Sense Voice 模型"""
self.recognizer = sherpa_onnx.OfflineRecognizer.from_sense_voice(
model=f"{self.model_path}/model.onnx",
tokens=f"{self.model_path}/tokens.txt",
num_threads = 2,
)
config = sherpa_onnx.VadModelConfig()
config.silero_vad.model = f"{self.model_path}/silero_vad.onnx"
config.silero_vad.threshold = 0.5
config.silero_vad.min_silence_duration = 0.1
config.silero_vad.min_speech_duration = 0.25
config.silero_vad.max_speech_duration = 8
config.sample_rate = 16000
self.window_size = config.silero_vad.window_size
self.vad = sherpa_onnx.VoiceActivityDetector(config, buffer_size_in_seconds=100)
self.buffer = []
self.offset = 0
self.started = False
self.started_time = .0
self.time_str = datetime.now().strftime('%H:%M:%S.%f')[:-3]
stdout_cmd('info', 'Shepra ONNX Sense Voice recognizer started.')
def send_audio_frame(self, data: bytes):
"""
发送音频帧给 SOSV 引擎，引擎将自动识别并将识别结果输出到标准输出中
Args:
data: 音频帧数据，采样率必须为 16000Hz
"""
caption = {}
caption['command'] = 'caption'
caption['translation'] = ''
data_np = np.frombuffer(data, dtype=np.int16).astype(np.float32)
self.buffer = np.concatenate([self.buffer, data_np])
while self.offset + self.window_size < len(self.buffer):
self.vad.accept_waveform(self.buffer[self.offset: self.offset + self.window_size])
if not self.started and self.vad.is_speech_detected():
self.started = True
self.started_time = time.time()
self.offset += self.window_size
if not self.started:
if len(self.buffer) > 10 * self.window_size:
self.offset -= len(self.buffer) - 10 * self.window_size
self.buffer = self.buffer[-10 * self.window_size:]
if self.started and time.time() - self.started_time > 0.2:
stream = self.recognizer.create_stream()
stream.accept_waveform(16000, self.buffer)
self.recognizer.decode_stream(stream)
text = stream.result.text.strip()
if text and self.prev_content != text:
caption['index'] = self.cur_id
caption['text'] = text
caption['time_s'] = self.time_str
caption['time_t'] = datetime.now().strftime('%H:%M:%S.%f')[:-3]
self.prev_content = text
stdout_obj(caption)
self.started_time = time.time()
while not self.vad.empty():
stream = self.recognizer.create_stream()
stream.accept_waveform(16000, self.vad.front.samples)
self.vad.pop()
self.recognizer.decode_stream(stream)
text = stream.result.text.strip()
caption['index'] = self.cur_id
caption['text'] = text
caption['time_s'] = self.time_str
caption['time_t'] = datetime.now().strftime('%H:%M:%S.%f')[:-3]
self.prev_content = ''
stdout_obj(caption)
self.cur_id += 1
self.time_str = datetime.now().strftime('%H:%M:%S.%f')[:-3]
self.buffer = []
self.offset = 0
self.started = False
self.started_time = .0
def translate(self):
"""持续读取共享数据中的音频帧，并进行语音识别，将识别结果输出到标准输出中"""
global shared_data
while shared_data.status == 'running':
chunk = shared_data.chunk_queue.get()
self.send_audio_frame(chunk)
def stop(self):
"""停止 Sense Voice 模型"""
stdout_cmd('info', 'Shepra ONNX Sense Voice recognizer closed.')

8
engine/audio2text/vosk.py
View File

@@ -4,6 +4,7 @@ import time
from datetime import datetime
from vosk import Model, KaldiRecognizer, SetLogLevel
from utils import shared_data
from utils import stdout_cmd, stdout_obj, google_translate, ollama_translate
@@ -82,6 +83,13 @@ class VoskRecognizer:
stdout_obj(caption)
def translate(self):
"""持续读取共享数据中的音频帧，并进行语音识别，将识别结果输出到标准输出中"""
global shared_data
while shared_data.status == 'running':
chunk = shared_data.chunk_queue.get()
self.send_audio_frame(chunk)
def stop(self):
"""停止 Vosk 引擎"""
stdout_cmd('info', 'Vosk recognizer closed.')

166
engine/main.py
View File

@@ -1,70 +1,120 @@
import wave
import argparse
from utils import stdout_cmd, stdout_err
from utils import thread_data, start_server
import threading
from utils import stdout, stdout_cmd
from utils import shared_data, start_server
from utils import merge_chunk_channels, resample_chunk_mono
from audio2text import InvalidParameter, GummyRecognizer
from audio2text import GummyRecognizer
from audio2text import VoskRecognizer
from audio2text import SosvRecognizer
from sysaudio import AudioStream
def audio_recording(stream: AudioStream, resample: bool, save = False):
global shared_data
stream.open_stream()
if save:
wf = wave.open(f'record.wav', 'wb')
wf.setnchannels(1)
wf.setsampwidth(stream.SAMP_WIDTH)
wf.setframerate(16000)
while shared_data.status == 'running':
raw_chunk = stream.read_chunk()
if raw_chunk is None: continue
if resample:
chunk = resample_chunk_mono(raw_chunk, stream.CHANNELS, stream.RATE, 16000)
else:
chunk = merge_chunk_channels(raw_chunk, stream.CHANNELS)
shared_data.chunk_queue.put(chunk)
if save: wf.writeframes(chunk) # type: ignore
if save: wf.close() # type: ignore
stream.close_stream_signal()
def main_gummy(s: str, t: str, a: int, c: int, k: str):
global thread_data
"""
Parameters:
s: Source language
t: Target language
k: Aliyun Bailian API key
"""
stream = AudioStream(a, c)
if t == 'none':
engine = GummyRecognizer(stream.RATE, s, None, k)
else:
engine = GummyRecognizer(stream.RATE, s, t, k)
stream.open_stream()
engine.start()
chunk_mono = bytes()
restart_count = 0
while thread_data.status == "running":
try:
chunk = stream.read_chunk()
if chunk is None: continue
chunk_mono = merge_chunk_channels(chunk, stream.CHANNELS)
try:
engine.send_audio_frame(chunk_mono)
except InvalidParameter as e:
restart_count += 1
if restart_count > 5:
stdout_err(str(e))
thread_data.status = "kill"
stdout_cmd('kill')
break
else:
stdout_cmd('info', f'Gummy engine stopped, restart attempt: {restart_count}...')
except KeyboardInterrupt:
break
engine.send_audio_frame(chunk_mono)
stream.close_stream()
stream_thread = threading.Thread(
target=audio_recording,
args=(stream, False),
daemon=True
)
stream_thread.start()
try:
engine.translate()
except KeyboardInterrupt:
stdout("Keyboard interrupt detected. Exiting...")
engine.stop()
def main_vosk(a: int, c: int, m: str, t: str, tm: str, on: str):
global thread_data
def main_vosk(a: int, c: int, vosk: str, t: str, tm: str, omn: str):
"""
Parameters:
a: Audio source: 0 for output, 1 for input
c: Chunk number in 1 second
vosk: Vosk model path
t: Target language
tm: Translation model type, ollama or google
omn: Ollama model name
"""
stream = AudioStream(a, c)
engine = VoskRecognizer(
m, None if t == 'none' else t,
tm, on
)
if t == 'none':
engine = VoskRecognizer(vosk, None, tm, omn)
else:
engine = VoskRecognizer(vosk, t, tm, omn)
stream.open_stream()
engine.start()
stream_thread = threading.Thread(
target=audio_recording,
args=(stream, True),
daemon=True
)
stream_thread.start()
try:
engine.translate()
except KeyboardInterrupt:
stdout("Keyboard interrupt detected. Exiting...")
engine.stop()
while thread_data.status == "running":
try:
chunk = stream.read_chunk()
if chunk is None: continue
chunk_mono = resample_chunk_mono(chunk, stream.CHANNELS, stream.RATE, 16000)
engine.send_audio_frame(chunk_mono)
except KeyboardInterrupt:
break
stream.close_stream()
def main_sosv(a: int, c: int, sosv: str, t: str, tm: str, omn: str):
"""
Parameters:
a: Audio source: 0 for output, 1 for input
c: Chunk number in 1 second
sosv: Sherpa-ONNX SenseVoice model path
t: Target language
tm: Translation model type, ollama or google
omn: Ollama model name
"""
stream = AudioStream(a, c)
if t == 'none':
engine = SosvRecognizer(sosv, None, tm, omn)
else:
engine = SosvRecognizer(sosv, t, tm, omn)
engine.start()
stream_thread = threading.Thread(
target=audio_recording,
args=(stream, True),
daemon=True
)
stream_thread.start()
try:
engine.translate()
except KeyboardInterrupt:
stdout("Keyboard interrupt detected. Exiting...")
engine.stop()
@@ -74,19 +124,22 @@ if __name__ == "__main__":
parser.add_argument('-e', '--caption_engine', default='gummy', help='Caption engine: gummy or vosk')
parser.add_argument('-a', '--audio_type', default=0, help='Audio stream source: 0 for output, 1 for input')
parser.add_argument('-c', '--chunk_rate', default=10, help='Number of audio stream chunks collected per second')
parser.add_argument('-p', '--port', default=8080, help='The port to run the server on, 0 for no server')
parser.add_argument('-p', '--port', default=0, help='The port to run the server on, 0 for no server')
parser.add_argument('-t', '--target_language', default='zh', help='Target language code, "none" for no translation')
# gummy only
parser.add_argument('-s', '--source_language', default='en', help='Source language code')
parser.add_argument('-k', '--api_key', default='', help='API KEY for Gummy model')
# vosk and sosv
parser.add_argument('-tm', '--translation_model', default='ollama', help='Model for translation: ollama or google')
parser.add_argument('-omn', '--ollama_name', default='', help='Ollama model name for translation')
# vosk only
parser.add_argument('-m', '--model_path', default='', help='The path to the vosk model.')
parser.add_argument('-tm', '--translation_model', default='', help='Google translate API KEY')
parser.add_argument('-on', '--ollama_name', default='', help='Ollama model name for translation')
parser.add_argument('-vosk', '--vosk_model', default='', help='The path to the vosk model.')
# sosv only
parser.add_argument('-sosv', '--sosv_model', default=None, help='The SenseVoice model path')
args = parser.parse_args()
if int(args.port) == 0:
thread_data.status = "running"
shared_data.status = "running"
else:
start_server(int(args.port))
@@ -102,7 +155,16 @@ if __name__ == "__main__":
main_vosk(
int(args.audio_type),
int(args.chunk_rate),
args.model_path,
args.vosk_model,
args.target_language,
args.translation_model,
args.ollama_name
)
elif args.caption_engine == 'sosv':
main_sosv(
int(args.audio_type),
int(args.chunk_rate),
args.sosv_model,
args.target_language,
args.translation_model,
args.ollama_name
@@ -110,5 +172,5 @@ if __name__ == "__main__":
else:
raise ValueError('Invalid caption engine specified.')
if thread_data.status == "kill":
if shared_data.status == "kill":
stdout_cmd('kill')

40
engine/sysaudio/darwin.py
View File

@@ -37,14 +37,13 @@ class AudioStream:
self.FORMAT = pyaudio.paInt16
self.SAMP_WIDTH = pyaudio.get_sample_size(self.FORMAT)
self.CHANNELS = int(self.device["maxInputChannels"])
self.RATE = int(self.device["defaultSampleRate"])
self.CHUNK = self.RATE // chunk_rate
self.DEFAULT_RATE = int(self.device["defaultSampleRate"])
self.CHUNK_RATE = chunk_rate
def reset_chunk_size(self, chunk_size: int):
"""
重新设置音频块大小
"""
self.CHUNK = chunk_size
self.RATE = 16000
self.CHUNK = self.RATE // self.CHUNK_RATE
self.open_stream()
self.close_stream()
def get_info(self):
dev_info = f"""
@@ -72,16 +71,27 @@ class AudioStream:
打开并返回系统音频输出流
"""
if self.stream: return self.stream
self.stream = self.mic.open(
format = self.FORMAT,
channels = int(self.CHANNELS),
rate = self.RATE,
input = True,
input_device_index = int(self.INDEX)
)
try:
self.stream = self.mic.open(
format = self.FORMAT,
channels = int(self.CHANNELS),
rate = self.RATE,
input = True,
input_device_index = int(self.INDEX)
)
except OSError:
self.RATE = self.DEFAULT_RATE
self.CHUNK = self.RATE // self.CHUNK_RATE
self.stream = self.mic.open(
format = self.FORMAT,
channels = int(self.CHANNELS),
rate = self.RATE,
input = True,
input_device_index = int(self.INDEX)
)
return self.stream
def read_chunk(self):
def read_chunk(self) -> bytes | None:
"""
读取音频数据
"""

11
engine/sysaudio/linux.py
View File

@@ -55,15 +55,10 @@ class AudioStream:
self.FORMAT = 16
self.SAMP_WIDTH = 2
self.CHANNELS = 2
self.RATE = 48000
self.RATE = 16000
self.CHUNK_RATE = chunk_rate
self.CHUNK = self.RATE // chunk_rate
def reset_chunk_size(self, chunk_size: int):
"""
重新设置音频块大小
"""
self.CHUNK = chunk_size
def get_info(self):
dev_info = f"""
音频捕获进程：
@@ -84,7 +79,7 @@ class AudioStream:
启动音频捕获进程
"""
self.process = subprocess.Popen(
["parec", "-d", self.source, "--format=s16le", "--rate=48000", "--channels=2"],
["parec", "-d", self.source, "--format=s16le", "--rate=16000", "--channels=2"],
stdout=subprocess.PIPE
)

38
engine/sysaudio/win.py
View File

@@ -61,14 +61,13 @@ class AudioStream:
self.FORMAT = pyaudio.paInt16
self.SAMP_WIDTH = pyaudio.get_sample_size(self.FORMAT)
self.CHANNELS = int(self.device["maxInputChannels"])
self.RATE = int(self.device["defaultSampleRate"])
self.CHUNK = self.RATE // chunk_rate
self.DEFAULT_RATE = int(self.device["defaultSampleRate"])
self.CHUNK_RATE = chunk_rate
def reset_chunk_size(self, chunk_size: int):
"""
重新设置音频块大小
"""
self.CHUNK = chunk_size
self.RATE = 16000
self.CHUNK = self.RATE // self.CHUNK_RATE
self.open_stream()
self.close_stream()
def get_info(self):
dev_info = f"""
@@ -96,13 +95,24 @@ class AudioStream:
打开并返回系统音频输出流
"""
if self.stream: return self.stream
self.stream = self.mic.open(
format = self.FORMAT,
channels = self.CHANNELS,
rate = self.RATE,
input = True,
input_device_index = self.INDEX
)
try:
self.stream = self.mic.open(
format = self.FORMAT,
channels = self.CHANNELS,
rate = self.RATE,
input = True,
input_device_index = self.INDEX
)
except OSError:
self.RATE = self.DEFAULT_RATE
self.CHUNK = self.RATE // self.CHUNK_RATE
self.stream = self.mic.open(
format = self.FORMAT,
channels = self.CHANNELS,
rate = self.RATE,
input = True,
input_device_index = self.INDEX
)
return self.stream
def read_chunk(self) -> bytes | None:

2
engine/utils/__init__.py
View File

@@ -5,6 +5,6 @@ from .audioprcs import (
resample_mono_chunk
)
from .sysout import stdout, stdout_err, stdout_cmd, stdout_obj, stderr
from .thdata import thread_data
from .shared import shared_data
from .server import start_server
from .translation import ollama_translate, google_translate

25
engine/utils/audioprcs.py
View File

@@ -49,9 +49,18 @@ def resample_chunk_mono(chunk: bytes, channels: int, orig_sr: int, target_sr: in
# (length,)
chunk_mono = np.mean(chunk_np.astype(np.float32), axis=1)
if orig_sr == target_sr:
return chunk_mono.astype(np.int16).tobytes()
ratio = target_sr / orig_sr
chunk_mono_r = samplerate.resample(chunk_mono, ratio, converter_type=mode)
chunk_mono_r = np.round(chunk_mono_r).astype(np.int16)
real_len = round(chunk_mono.shape[0] * ratio)
if(chunk_mono_r.shape[0] > real_len):
chunk_mono_r = chunk_mono_r[:real_len]
else:
while chunk_mono_r.shape[0] < real_len:
chunk_mono_r = np.append(chunk_mono_r, chunk_mono_r[-1])
return chunk_mono_r.tobytes()
@@ -81,9 +90,18 @@ def resample_chunk_mono_np(chunk: bytes, channels: int, orig_sr: int, target_sr:
# (length,)
chunk_mono = np.mean(chunk_np.astype(np.float32), axis=1)
if orig_sr == target_sr:
return chunk_mono.astype(dtype)
ratio = target_sr / orig_sr
chunk_mono_r = samplerate.resample(chunk_mono, ratio, converter_type=mode)
chunk_mono_r = chunk_mono_r.astype(dtype)
real_len = round(chunk_mono.shape[0] * ratio)
if(chunk_mono_r.shape[0] > real_len):
chunk_mono_r = chunk_mono_r[:real_len]
else:
while chunk_mono_r.shape[0] < real_len:
chunk_mono_r = np.append(chunk_mono_r, chunk_mono_r[-1])
return chunk_mono_r
@@ -100,9 +118,16 @@ def resample_mono_chunk(chunk: bytes, orig_sr: int, target_sr: int, mode="sinc_b
Return:
单通道音频数据块
"""
if orig_sr == target_sr: return chunk
chunk_np = np.frombuffer(chunk, dtype=np.int16)
chunk_np = chunk_np.astype(np.float32)
ratio = target_sr / orig_sr
chunk_r = samplerate.resample(chunk_np, ratio, converter_type=mode)
chunk_r = np.round(chunk_r).astype(np.int16)
real_len = round(chunk_np.shape[0] * ratio)
if(chunk_r.shape[0] > real_len):
chunk_r = chunk_r[:real_len]
else:
while chunk_r.shape[0] < real_len:
chunk_r = np.append(chunk_r, chunk_r[-1])
return chunk_r.tobytes()

12
engine/utils/server.py
View File

@@ -1,13 +1,12 @@
import socket
import threading
import json
# import time
from utils import thread_data, stdout_cmd, stderr
from utils import shared_data, stdout_cmd, stderr
def handle_client(client_socket):
global thread_data
while thread_data.status == 'running':
global shared_data
while shared_data.status == 'running':
try:
data = client_socket.recv(4096).decode('utf-8')
if not data:
@@ -15,13 +14,13 @@ def handle_client(client_socket):
data = json.loads(data)
if data['command'] == 'stop':
thread_data.status = 'stop'
shared_data.status = 'stop'
break
except Exception as e:
stderr(f'Communication error: {e}')
break
thread_data.status = 'stop'
shared_data.status = 'stop'
client_socket.close()
@@ -34,7 +33,6 @@ def start_server(port: int):
stderr(str(e))
stdout_cmd('kill')
return
# time.sleep(20)
stdout_cmd('connect')
client, addr = server.accept()

8
engine/utils/shared.py Normal file
View File

@@ -0,0 +1,8 @@
import queue
class SharedData:
def __init__(self):
self.status = "running"
self.chunk_queue = queue.Queue()
shared_data = SharedData()

5
engine/utils/thdata.py
View File

@@ -1,5 +0,0 @@
class ThreadData:
def __init__(self):
self.status = "running"
thread_data = ThreadData()

4
src/main/utils/CaptionEngine.ts
View File

@@ -81,9 +81,9 @@ export class CaptionEngine {
}
else if(allConfig.controls.engine === 'vosk'){
this.command.push('-e', 'vosk')
this.command.push('-m', `"${allConfig.controls.modelPath}"`)
this.command.push('-vosk', `"${allConfig.controls.modelPath}"`)
this.command.push('-tm', allConfig.controls.transModel)
this.command.push('-on', allConfig.controls.ollamaName)
this.command.push('-omn', allConfig.controls.ollamaName)
}
}
Log.info('Engine Path:', this.appPath)