update add fix

.gitignore

@@ -68,7 +68,7 @@ instance/
 
 # Sphinx documentation
 docs/_build/
-
+test/
 # PyBuilder
 .pybuilder/
 target/

moyoyo_asr_models/Qwen2.5-3B.Q5_0.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:35a1eec88348fd963b06bb2e3ba8d4d27d4f7a79d9ed4bab645f931d7b2b7ac7
+size 2169666016

moyoyo_asr_models/qwen2.5-7b-instruct-q5_0-00001-of-00002.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:715a0c76b1b074ed65fe0fabf88b1df6f5f638318f15ba9f62a04d493e236d5f
+size 4001112160

moyoyo_asr_models/qwen2.5-7b-instruct-q5_0-00002-of-00002.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6a254c888ebffcddf98b0c277433a339d9a8cbf970f246b0b6e6a9195e5d6cfb
+size 1314064416

pyproject.toml

@@ -6,12 +6,15 @@ readme = "README.md"
 requires-python = ">=3.11"
 dependencies = [
     "av>=14.2.0",
+    "fastapi>=0.115.12",
     "librosa>=0.11.0",
     "numpy>=2.1.3",
     "onnxruntime>=1.21.0",
     "pyaudio>=0.2.14",
     "pydantic>=2.11.2",
+    "pydub>=0.25.1",
     "setuptools>=78.1.0",
+    "silero-vad>=5.1.2",
     "soundfile>=0.13.1",
     "torch>=2.6.0",
     "tqdm>=4.67.1",

requirements.txt

@@ -2,6 +2,8 @@
 #    uv pip compile pyproject.toml -o requirements.txt
 annotated-types==0.7.0
     # via pydantic
+anyio==4.9.0
+    # via starlette
 audioread==3.0.1
     # via librosa
 av==14.3.0
@@ -16,8 +18,8 @@ coloredlogs==15.0.1
     # via onnxruntime
 decorator==5.2.1
     # via librosa
-diskcache==5.6.3
-    # via llama-cpp-python
+fastapi==0.115.12
+    # via trans (pyproject.toml)
 filelock==3.18.0
     # via torch
 flatbuffers==25.2.10
@@ -27,11 +29,11 @@ fsspec==2025.3.2
 humanfriendly==10.0
     # via coloredlogs
 idna==3.10
-    # via requests
-jinja2==3.1.6
     # via
-    #   llama-cpp-python
-    #   torch
+    #   anyio
+    #   requests
+jinja2==3.1.6
+    # via torch
 joblib==1.4.2
     # via
     #   librosa
@@ -40,8 +42,6 @@ lazy-loader==0.4
     # via librosa
 librosa==0.11.0
     # via trans (pyproject.toml)
-llama-cpp-python==0.3.8
-    # via trans (pyproject.toml)
 llvmlite==0.44.0
     # via numba
 markupsafe==3.0.2
@@ -58,25 +58,23 @@ numpy==2.1.3
     # via
     #   trans (pyproject.toml)
     #   librosa
-    #   llama-cpp-python
     #   numba
     #   onnxruntime
-    #   pywhispercpp
     #   scikit-learn
     #   scipy
     #   soundfile
     #   soxr
 onnxruntime==1.21.0
-    # via trans (pyproject.toml)
+    # via
+    #   trans (pyproject.toml)
+    #   silero-vad
 packaging==24.2
     # via
     #   lazy-loader
     #   onnxruntime
     #   pooch
 platformdirs==4.3.7
-    # via
-    #   pooch
-    #   pywhispercpp
+    # via pooch
 pooch==1.8.2
     # via librosa
 protobuf==6.30.2
@@ -86,15 +84,15 @@ pyaudio==0.2.14
 pycparser==2.22
     # via cffi
 pydantic==2.11.2
-    # via trans (pyproject.toml)
+    # via
+    #   trans (pyproject.toml)
+    #   fastapi
 pydantic-core==2.33.1
     # via pydantic
-pywhispercpp==1.3.0
+pydub==0.25.1
     # via trans (pyproject.toml)
 requests==2.32.3
-    # via
-    #   pooch
-    #   pywhispercpp
+    # via pooch
 scikit-learn==1.6.1
     # via librosa
 scipy==1.15.2
@@ -103,12 +101,18 @@ scipy==1.15.2
     #   scikit-learn
 setuptools==78.1.0
     # via trans (pyproject.toml)
+silero-vad==5.1.2
+    # via trans (pyproject.toml)
+sniffio==1.3.1
+    # via anyio
 soundfile==0.13.1
     # via
     #   trans (pyproject.toml)
     #   librosa
 soxr==0.5.0.post1
     # via librosa
+starlette==0.46.1
+    # via fastapi
 sympy==1.13.1
     # via
     #   onnxruntime
@@ -116,15 +120,19 @@ sympy==1.13.1
 threadpoolctl==3.6.0
     # via scikit-learn
 torch==2.6.0
-    # via trans (pyproject.toml)
-tqdm==4.67.1
     # via
     #   trans (pyproject.toml)
-    #   pywhispercpp
+    #   silero-vad
+    #   torchaudio
+torchaudio==2.6.0
+    # via silero-vad
+tqdm==4.67.1
+    # via trans (pyproject.toml)
 typing-extensions==4.13.1
     # via
+    #   anyio
+    #   fastapi
     #   librosa
-    #   llama-cpp-python
     #   pydantic
     #   pydantic-core
     #   torch

transcribe/helpers/vadprocessor.py

@@ -0,0 +1,251 @@
+import gc
+import time
+import numpy as np
+import onnxruntime
+from datetime import timedelta
+from pydub import AudioSegment
+from silero_vad import load_silero_vad, get_speech_timestamps
+import os
+import logging
+
+
+class SileroVADProcessor:
+    """
+    A class for processing audio files using Silero VAD to detect voice activity
+    and extract voice segments from audio files.
+    """
+
+    def __init__(self,
+                 activate_threshold=0.5,
+                 fusion_threshold=0.3,
+                 min_speech_duration=0.25,
+                 max_speech_duration=20,
+                 min_silence_duration=250,
+                 sample_rate=16000,
+                 ort_providers=None):
+        """
+        Initialize the SileroVADProcessor.
+
+        Args:
+            activate_threshold (float): Threshold for voice activity detection
+            fusion_threshold (float): Threshold for merging close speech segments (seconds)
+            min_speech_duration (float): Minimum duration of speech to be considered valid (seconds)
+            max_speech_duration (float): Maximum duration of speech (seconds)
+            min_silence_duration (int): Minimum silence duration (ms)
+            sample_rate (int): Sample rate of the audio (8000 or 16000 Hz)
+            ort_providers (list): ONNX Runtime providers for acceleration
+        """
+        # VAD parameters
+        self.activate_threshold = activate_threshold
+        self.fusion_threshold = fusion_threshold
+        self.min_speech_duration = min_speech_duration
+        self.max_speech_duration = max_speech_duration
+        self.min_silence_duration = min_silence_duration
+        self.sample_rate = sample_rate
+        self.ort_providers = ort_providers if ort_providers else []
+
+        # Initialize logger
+        self.logger = logging.getLogger(__name__)
+
+        # Load Silero VAD model
+        self._init_onnx_session()
+        self.silero_vad = load_silero_vad(onnx=True)
+
+    def _init_onnx_session(self):
+        """Initialize ONNX Runtime session with appropriate settings."""
+        session_opts = onnxruntime.SessionOptions()
+        session_opts.log_severity_level = 3
+        session_opts.inter_op_num_threads = 0
+        session_opts.intra_op_num_threads = 0
+        session_opts.enable_cpu_mem_arena = True
+        session_opts.execution_mode = onnxruntime.ExecutionMode.ORT_SEQUENTIAL
+        session_opts.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
+
+        session_opts.add_session_config_entry("session.intra_op.allow_spinning", "1")
+        session_opts.add_session_config_entry("session.inter_op.allow_spinning", "1")
+        session_opts.add_session_config_entry("session.set_denormal_as_zero", "1")
+
+        # Set the session_opts to be used by silero_vad
+        # onnxruntime.capi._pybind_state.get_default_session_options(session_opts)
+
+    def load_audio(self, audio_path):
+        """
+        Load audio file and prepare it for VAD processing.
+
+        Args:
+            audio_path (str): Path to the audio file
+
+        Returns:
+            numpy.ndarray: Audio data as numpy array
+        """
+        self.logger.info(f"Loading audio from {audio_path}")
+        audio_segment = AudioSegment.from_file(audio_path)
+        audio_segment = audio_segment.set_channels(1).set_frame_rate(self.sample_rate)
+
+        # Convert to numpy array and normalize
+        dtype = np.float16 if self.use_gpu_fp16 else np.float32
+        audio_array = np.array(audio_segment.get_array_of_samples(), dtype=dtype) * 0.000030517578  # 1/32768
+
+        self.audio_segment = audio_segment  # Store for later use
+        return audio_array
+    
+    @property
+    def model(self):
+        return self.silero_vad
+
+    def process_timestamps(self, timestamps):
+        """
+        Process VAD timestamps: filter short segments and merge close segments.
+
+        Args:
+            timestamps (list): List of (start, end) tuples
+
+        Returns:
+            list: Processed list of (start, end) tuples
+        """
+        # Filter out short durations
+        filtered_timestamps = [(start, end) for start, end in timestamps
+                               if (end - start) >= self.min_speech_duration]
+
+        # Fuse timestamps in two passes for better merging
+        fused_timestamps_1st = []
+        for start, end in filtered_timestamps:
+            if fused_timestamps_1st and (start - fused_timestamps_1st[-1][1] <= self.fusion_threshold):
+                fused_timestamps_1st[-1] = (fused_timestamps_1st[-1][0], end)
+            else:
+                fused_timestamps_1st.append((start, end))
+
+        fused_timestamps_2nd = []
+        for start, end in fused_timestamps_1st:
+            if fused_timestamps_2nd and (start - fused_timestamps_2nd[-1][1] <= self.fusion_threshold):
+                fused_timestamps_2nd[-1] = (fused_timestamps_2nd[-1][0], end)
+            else:
+                fused_timestamps_2nd.append((start, end))
+
+        return fused_timestamps_2nd
+
+    def format_time(self, seconds):
+        """
+        Convert seconds to VTT time format 'hh:mm:ss.mmm'.
+
+        Args:
+            seconds (float): Time in seconds
+
+        Returns:
+            str: Formatted time string
+        """
+        td = timedelta(seconds=seconds)
+        td_sec = td.total_seconds()
+        total_seconds = int(td_sec)
+        milliseconds = int((td_sec - total_seconds) * 1000)
+        hours = total_seconds // 3600
+        minutes = (total_seconds % 3600) // 60
+        seconds = total_seconds % 60
+        return f"{hours:02}:{minutes:02}:{seconds:02}.{milliseconds:03}"
+
+    def detect_speech(self, audio:np.array):
+        """
+        Run VAD on the audio file to detect speech segments.
+
+        Args:
+            audio_path (str): Path to the audio file
+
+        Returns:
+            list: List of processed timestamps as (start, end) tuples
+        """
+        self.logger.info("Starting VAD process")
+        start_time = time.time()
+        # Get speech timestamps
+        raw_timestamps = get_speech_timestamps(
+            audio,
+            model=self.silero_vad,
+            threshold=self.activate_threshold,
+            max_speech_duration_s=self.max_speech_duration,
+            min_speech_duration_ms=int(self.min_speech_duration * 1000),
+            min_silence_duration_ms=self.min_silence_duration,
+            return_seconds=True
+        )
+
+        # Convert to simple format and process
+        timestamps = [(item['start'], item['end']) for item in raw_timestamps]
+        processed_timestamps = self.process_timestamps(timestamps)
+
+        # Clean up
+        del audio
+        gc.collect()
+
+        self.logger.info(f"VAD completed in {time.time() - start_time:.3f} seconds")
+        return processed_timestamps
+
+        """
+        Save timestamps in both second and sample indices formats.
+
+        Args:
+            timestamps (list): List of (start, end) tuples
+            output_prefix (str): Prefix for output files
+        """
+        # Save timestamps in seconds (VTT format)
+        seconds_path = f"{output_prefix}_timestamps_second.txt"
+        with open(seconds_path, "w", encoding='UTF-8') as file:
+            self.logger.info("Saving timestamps in seconds format")
+            for start, end in timestamps:
+                s_time = self.format_time(start)
+                e_time = self.format_time(end)
+                line = f"{s_time} --> {e_time}\n"
+                file.write(line)
+
+        # Save timestamps in sample indices
+        indices_path = f"{output_prefix}_timestamps_indices.txt"
+        with open(indices_path, "w", encoding='UTF-8') as file:
+            self.logger.info("Saving timestamps in indices format")
+            for start, end in timestamps:
+                line = f"{int(start * self.sample_rate)} --> {int(end * self.sample_rate)}\n"
+                file.write(line)
+
+        self.logger.info(f"Timestamps saved to {seconds_path} and {indices_path}")
+
+    def extract_speech_segments(self, audio_segment, timestamps):
+        """
+        Extract speech segments from the audio and combine them into a single audio file.
+
+        Args:
+            timestamps (list): List of (start, end) tuples indicating speech segments
+
+        Returns:
+            AudioSegment: The combined speech segments
+        """
+        audio_segment = audio_segment.numpy()
+        combined_speech = np.array([], dtype=np.float32)
+
+        # Extract and combine each speech segment
+        for i, (start, end) in enumerate(timestamps):
+            # Convert seconds to milliseconds for pydub
+            start_ms = int(start * 1000)
+            end_ms = int(end * 1000)
+
+            # Ensure the end time does not exceed the length of the audio segment
+            if end_ms > len(audio_segment):
+                end_ms = len(audio_segment)
+
+            # Extract the segment
+            segment = audio_segment[start_ms:end_ms]
+
+            # Add to combined audio
+            combined_speech = np.append(combined_speech, segment)
+
+        return combined_speech
+
+    def process_audio(self, audio_array:np.array):
+        """
+        Complete processing pipeline: detect speech, save timestamps, and optionally extract speech.
+
+        Returns:
+            tuple: (timestamps, output_speech_path if extract_speech else None)
+        """
+
+        # Run VAD to detect speech
+        timestamps = self.detect_speech(audio_array)
+
+        combined_speech = self.extract_speech_segments(audio_array, timestamps)
+
+        return timestamps, combined_speech

transcribe/pipelines/__init__.py

@@ -1,4 +1,5 @@
 
 from .pipe_translate import TranslatePipe
 from .pipe_whisper import WhisperPipe
+from .pipe_vad import VadPipe
 from .base import MetaItem

transcribe/pipelines/base.py

@@ -12,6 +12,7 @@ class Segment:
 @dataclass
 class MetaItem:
     segments: list[Segment] = field(default_factory=list)
+    source_audio: bytes = b""
     audio: bytes = b''
     transcribe_content: str = ''
     translate_content: str = ''

transcribe/pipelines/pipe_translate.py

@@ -1,7 +1,7 @@
 
 from .base import MetaItem, BasePipe, Segment
 from llama_cpp import Llama
-from ..translator import QwenTranslator
+from ..helpers.translator import QwenTranslator
 from config import LLM_MODEL_PATH, LLM_SYS_PROMPT
 
 

transcribe/pipelines/pipe_vad.py

@@ -1,3 +1,70 @@
 
-
 from .base import MetaItem, BasePipe
+from ..helpers.vadprocessor import SileroVADProcessor
+import numpy as np
+from silero_vad import read_audio, get_speech_timestamps,collect_chunks, VADIterator
+import torch
+
+class VadPipe(BasePipe):
+    model = None
+    sample_rate=16000
+    window_size_samples = 512
+
+    @classmethod
+    def init(cls):
+        if cls.model is None:
+            cls.model = SileroVADProcessor(
+            activate_threshold=0.5,
+            fusion_threshold=0.3,
+            min_speech_duration=0.25,
+            max_speech_duration=20,
+            min_silence_duration=250,
+            sample_rate=cls.sample_rate
+        )
+
+    @property
+    def vad_iterator(self):
+        return VADIterator(self.model.silero_vad, sampling_rate=self.sample_rate,)
+
+    def process(self, in_data: MetaItem) -> MetaItem:
+        source_audio = in_data.source_audio
+        source_audio = np.frombuffer(source_audio, dtype=np.float32)
+        speech_segments = []
+        is_speech_active = False
+        # current_segment_end = len(source_audio)
+       
+        for i in range(0, len(source_audio), self.window_size_samples):
+
+            window = source_audio[i:i+self.window_size_samples]
+         
+            if len(window) < self.window_size_samples:
+                padded_window = np.zeros(self.window_size_samples, dtype=np.float32)
+                padded_window[:len(window)] = window
+                window = padded_window  
+
+            speech_dict = self.vad_iterator(window, return_seconds=False)
+            if not speech_dict:
+                continue
+   
+            # 计算当前偏移量
+
+            if speech_dict and 'start' in speech_dict and not is_speech_active:
+                is_speech_active = True
+                # current_segment_start = speech_dict['start'] + i
+            
+            if is_speech_active:
+                speech_segments.append(window)
+            # # 如果检测到语音结束
+            # if speech_dict and 'end' in speech_dict and is_speech_active:
+            #     # 调整语音结束时间，加上窗口偏移
+            #     current_segment_end = min(speech_dict['end'] + i, current_segment_end)
+            #     is_speech_active = False
+            #     speech_audio = source_audio[current_segment_start: current_segment_end]
+            #     speech_segments.append(speech_audio)
+
+        self.vad_iterator.reset_states()
+        combied_audio = np.concatenate(speech_segments, axis=0).tobytes() if len(speech_segments) else b""
+        in_data.audio = combied_audio
+        in_data.source_audio = b""
+        
+        return in_data

transcribe/pipelines/pipe_whisper.py

@@ -1,7 +1,7 @@
 
 
 from .base import MetaItem, BasePipe, Segment
-from ..whisper import WhisperCPP
+from ..helpers.whisper import WhisperCPP
 
 
 class WhisperPipe(BasePipe):

transcribe/translatepipes.py

@@ -1,4 +1,4 @@
-from transcribe.pipelines import WhisperPipe, TranslatePipe, MetaItem
+from transcribe.pipelines import WhisperPipe, TranslatePipe, MetaItem, VadPipe
 import multiprocessing  as mp
 import config
 
@@ -10,21 +10,23 @@ class TranslatePipes:
         # self.result_queue = mp.Queue()
         
         # whisper 转录
-        self._whisper_pipe = WhisperPipe()
+        self._whisper_pipe = self._launch_process(WhisperPipe())
         
         # llm 翻译
-        self._translate_pipe = TranslatePipe()
+        self._translate_pipe = self._launch_process(TranslatePipe())
 
-        
-        self._whisper_pipe.daemon = True
-        self._whisper_pipe.start()
-
-        self._translate_pipe.daemon = True
-        self._translate_pipe.start()
+        # vad 
+        self._vad_pipe = self._launch_process(VadPipe())
+    
+    def _launch_process(self, process_obj):
+        process_obj.daemon = True
+        process_obj.start()
+        return process_obj
 
     def wait_ready(self):
         self._whisper_pipe.wait()
         self._translate_pipe.wait()
+        self._vad_pipe.wait()
     
     def translate(self, text, src_lang, dst_lang) -> MetaItem:
         item = MetaItem(
@@ -40,13 +42,20 @@ class TranslatePipes:
         self._whisper_pipe.input_queue.put(item)
         return self._whisper_pipe.output_queue.get()
     
+    def voice_detect(self, audio_buffer:bytes) -> MetaItem:
+        item = MetaItem(source_audio=audio_buffer)
+        self._vad_pipe.input_queue.put(item)
+        return self._vad_pipe.output_queue.get()
+
+    
 
 if __name__ == "__main__":
     import soundfile
     tp = TranslatePipes()
     # result = tp.translate("你好，今天天气怎么样?", src_lang="zh", dst_lang="en")
     mel, _, = soundfile.read("assets/jfk.flac")
-    result = tp.transcrible(mel, 'en')
+    # result = tp.transcrible(mel, 'en')
+    result = tp.voice_detect(mel)
     print(result)
 
 

transcribe/utils.py

@@ -2,6 +2,8 @@ import os
 import textwrap
 from pathlib import Path
 import logging
+import numpy as np
+from scipy.io.wavfile import write
 
 
 import av
@@ -86,3 +88,7 @@ def resample(file: str, sr: int = 16000):
 
     output_container.close()
     return resampled_file
+
+
+def save_to_wave(filename, data:np.ndarray, sample_rate=16000):
+    write(filename, sample_rate, data)

transcribe/whisper_llm_serve.py

@@ -5,13 +5,13 @@ import multiprocessing as mp
 import numpy as np
 from logging import getLogger
 
-import config
+from .utils import save_to_wave
 import time
 import json
 import threading
 from .server import ServeClientBase
-
-from scipy.io.wavfile import write
+import queue
+import collections
 from api_model import TransResult, Message
 from .utils import log_block
 from .translatepipes import TranslatePipes
@@ -23,14 +23,11 @@ translate_pipes = TranslatePipes()
 translate_pipes.wait_ready()
 logger.info("Pipeline is ready.")
 
-def save_to_wave(filename, data:np.ndarray, sample_rate=16000):
-    write(filename, sample_rate, data)
 
 
 
 class PyWhiperCppServe(ServeClientBase):
    
-    
     def __init__(self, websocket, language=None, dst_lang=None, client_uid=None,):
         super().__init__(client_uid, websocket)
         self.language = language
@@ -42,16 +39,23 @@ class PyWhiperCppServe(ServeClientBase):
         
         self.lock = threading.Lock()
         self.frames_np = None
+        self._frame_queue = queue.Queue()
+        self._previous_frame_queue = collections.deque(maxlen=2)
         self.sample_rate = 16000
  
         self.send_ready_state()
         
-        self.trans_thread = threading.Thread(target=self.speech_to_text)
-        self.trans_thread.daemon = True
-        self.trans_thread.start()
+        self.run_in_thread(self.speech_to_text)
+        self.run_in_thread(self.get_frame_from_queue)
+
+        self.text_sep = "" if self.language == "zh" else " "
+
+    def run_in_thread(self, func):
+        t = threading.Thread(target=func)
+        t.daemon = True
+        t.start()
 
     def send_ready_state(self):
-  
         self.websocket.send(json.dumps({
             "uid": self.client_uid,
             "message": self.SERVER_READY,
@@ -63,12 +67,34 @@ class PyWhiperCppServe(ServeClientBase):
         self.dst_lang = dst_lang
 
     def add_frames(self, frame_np):
-        # self._audio_queue.put(frame_np)
-        with self.lock:
-            if self.frames_np is None:
-                self.frames_np = frame_np.copy()
-            else:
-                self.frames_np = np.append(self.frames_np,frame_np)
+        self._frame_queue.put(frame_np)
+
+    def get_prev_frame(self, ):
+        if len(self._previous_frame_queue) == 2:
+            return self._previous_frame_queue[-1]
+
+    
+    def get_frame_from_queue(self,):
+        while True:
+            try:
+                frame_np = self._frame_queue.get(timeout=0.1)
+                # frame_np = item.source_audio
+                # self._previous_frame_queue.appendleft(frame_np.copy())
+                # prev_frame_np = self.get_prev_frame()
+                # if prev_frame_np is not None:
+                #     frame_np = np.concatenate([prev_frame_np[int(-0.05 * self.sample_rate):],frame_np], axis=0)
+                # item = translate_pipes.voice_detect(frame_np.tobytes())
+                # if item.audio == b"":
+                #     continue 
+                # frame_np = np.frombuffer(item.audio, dtype=np.float32)
+                with self.lock:
+                    if self.frames_np is None:
+                        self.frames_np = frame_np.copy()
+                    else:
+                        self.frames_np = np.append(self.frames_np,frame_np)
+            except queue.Empty:
+                pass
+          
 
 
     def update_audio_buffer(self, last_offset):
@@ -99,8 +125,10 @@ class PyWhiperCppServe(ServeClientBase):
         log_block("LLM translate input", f"{text}")
         start_time = time.perf_counter()
         ret =  translate_pipes.translate(text, self.language, self.dst_lang)
+        translated_text = ret.translate_content
         log_block("LLM translate time", f"{(time.perf_counter() - start_time):.3f}", "s")
-        return ret.translate_content
+        log_block("LLM translate out", f"{translated_text}")
+        return translated_text
     
 
     
@@ -109,8 +137,8 @@ class PyWhiperCppServe(ServeClientBase):
         #   当左边确认后，右边段才会进入观察
         #   当左边确认后，会从缓冲区中删除对应的buffer，减少下次输入的数据量
         left_watch_idx, left_watch_sequences, right_watch_sequences, is_end_sentence = segments_split(segments, audio_buffer)
-        left_watch_string = "".join(i.text for i in left_watch_sequences)
-        right_watch_string = "".join(i.text for i in right_watch_sequences)
+        left_watch_string = self.text_sep.join(i.text for i in left_watch_sequences)
+        right_watch_string = self.text_sep.join(i.text for i in right_watch_sequences)
         
         if left_watch_idx != 0:
             # 将观察字符串临时存储
@@ -121,8 +149,8 @@ class PyWhiperCppServe(ServeClientBase):
         
         # 整句消除 后两句之前的内容
         left_watch_idx, left_watch_sequences, right_watch_sequences, is_end_sentence = sequences_split(segments, audio_buffer)
-        left_watch_string = "".join(i.text for i in left_watch_sequences)
-        right_watch_string = "".join(i.text for i in right_watch_sequences)
+        left_watch_string = self.text_sep.join(i.text for i in left_watch_sequences)
+        right_watch_string = self.text_sep.join(i.text for i in right_watch_sequences)
         if left_watch_idx != 0:
             return left_watch_idx, left_watch_string, right_watch_string, is_end_sentence
         
@@ -157,7 +185,7 @@ class PyWhiperCppServe(ServeClientBase):
             #     logger.error(f"{e}")
     
     def handle_transcription_output(self, segments, audio_buffer):
-        texts  = "".join(i.text for i in segments)
+        texts  = self.text_sep.join(i.text for i in segments)
         if not len(texts):
             return
         self._segment_manager.handle(texts)

uv.lock

@@ -16,6 +16,20 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/78/b6/6307fbef88d9b5ee7421e68d78a9f162e0da4900bc5f5793f6d3d0e34fb8/annotated_types-0.7.0-py3-none-any.whl", hash = "sha256:1f02e8b43a8fbbc3f3e0d4f0f4bfc8131bcb4eebe8849b8e5c773f3a1c582a53", size = 13643 },
 ]
 
+[[package]]
+name = "anyio"
+version = "4.9.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "idna" },
+    { name = "sniffio" },
+    { name = "typing-extensions", marker = "python_full_version < '3.13'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/95/7d/4c1bd541d4dffa1b52bd83fb8527089e097a106fc90b467a7313b105f840/anyio-4.9.0.tar.gz", hash = "sha256:673c0c244e15788651a4ff38710fea9675823028a6f08a5eda409e0c9840a028", size = 190949 }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/a1/ee/48ca1a7c89ffec8b6a0c5d02b89c305671d5ffd8d3c94acf8b8c408575bb/anyio-4.9.0-py3-none-any.whl", hash = "sha256:9f76d541cad6e36af7beb62e978876f3b41e3e04f2c1fbf0884604c0a9c4d93c", size = 100916 },
+]
+
 [[package]]
 name = "audioop-lts"
 version = "0.2.1"
@@ -223,6 +237,20 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/4e/8c/f3147f5c4b73e7550fe5f9352eaa956ae838d5c51eb58e7a25b9f3e2643b/decorator-5.2.1-py3-none-any.whl", hash = "sha256:d316bb415a2d9e2d2b3abcc4084c6502fc09240e292cd76a76afc106a1c8e04a", size = 9190 },
 ]
 
+[[package]]
+name = "fastapi"
+version = "0.115.12"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "pydantic" },
+    { name = "starlette" },
+    { name = "typing-extensions" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/f4/55/ae499352d82338331ca1e28c7f4a63bfd09479b16395dce38cf50a39e2c2/fastapi-0.115.12.tar.gz", hash = "sha256:1e2c2a2646905f9e83d32f04a3f86aff4a286669c6c950ca95b5fd68c2602681", size = 295236 }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/50/b3/b51f09c2ba432a576fe63758bddc81f78f0c6309d9e5c10d194313bf021e/fastapi-0.115.12-py3-none-any.whl", hash = "sha256:e94613d6c05e27be7ffebdd6ea5f388112e5e430c8f7d6494a9d1d88d43e814d", size = 95164 },
+]
+
 [[package]]
 name = "filelock"
 version = "3.18.0"
@@ -831,6 +859,15 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/12/6f/5596dc418f2e292ffc661d21931ab34591952e2843e7168ea5a52591f6ff/pydantic_core-2.33.1-pp311-pypy311_pp73-win_amd64.whl", hash = "sha256:f995719707e0e29f0f41a8aa3bcea6e761a36c9136104d3189eafb83f5cec5e5", size = 2080951 },
 ]
 
+[[package]]
+name = "pydub"
+version = "0.25.1"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/fe/9a/e6bca0eed82db26562c73b5076539a4a08d3cffd19c3cc5913a3e61145fd/pydub-0.25.1.tar.gz", hash = "sha256:980a33ce9949cab2a569606b65674d748ecbca4f0796887fd6f46173a7b0d30f", size = 38326 }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/a6/53/d78dc063216e62fc55f6b2eebb447f6a4b0a59f55c8406376f76bf959b08/pydub-0.25.1-py2.py3-none-any.whl", hash = "sha256:65617e33033874b59d87db603aa1ed450633288aefead953b30bded59cb599a6", size = 32327 },
+]
+
 [[package]]
 name = "pyreadline3"
 version = "3.5.4"
@@ -944,6 +981,29 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/54/21/f43f0a1fa8b06b32812e0975981f4677d28e0f3271601dc88ac5a5b83220/setuptools-78.1.0-py3-none-any.whl", hash = "sha256:3e386e96793c8702ae83d17b853fb93d3e09ef82ec62722e61da5cd22376dcd8", size = 1256108 },
 ]
 
+[[package]]
+name = "silero-vad"
+version = "5.1.2"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "onnxruntime" },
+    { name = "torch" },
+    { name = "torchaudio" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/b1/b4/d0311b2e6220a11f8f4699f4a278cb088131573286cdfe804c87c7eb5123/silero_vad-5.1.2.tar.gz", hash = "sha256:c442971160026d2d7aa0ad83f0c7ee86c89797a65289fe625c8ea59fc6fb828d", size = 5098526 }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/98/f7/5ae11d13fbb733cd3bfd7ff1c3a3902e6f55437df4b72307c1f168146268/silero_vad-5.1.2-py3-none-any.whl", hash = "sha256:93b41953d7774b165407fda6b533c119c5803864e367d5034dc626c82cfdf661", size = 5026737 },
+]
+
+[[package]]
+name = "sniffio"
+version = "1.3.1"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/a2/87/a6771e1546d97e7e041b6ae58d80074f81b7d5121207425c964ddf5cfdbd/sniffio-1.3.1.tar.gz", hash = "sha256:f4324edc670a0f49750a81b895f35c3adb843cca46f0530f79fc1babb23789dc", size = 20372 }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/e9/44/75a9c9421471a6c4805dbf2356f7c181a29c1879239abab1ea2cc8f38b40/sniffio-1.3.1-py3-none-any.whl", hash = "sha256:2f6da418d1f1e0fddd844478f41680e794e6051915791a034ff65e5f100525a2", size = 10235 },
+]
+
 [[package]]
 name = "soundfile"
 version = "0.13.1"
@@ -1018,6 +1078,18 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/34/ae/e3707f6c1bc6f7aa0df600ba8075bfb8a19252140cd595335be60e25f9ee/standard_sunau-3.13.0-py3-none-any.whl", hash = "sha256:53af624a9529c41062f4c2fd33837f297f3baa196b0cfceffea6555654602622", size = 7364 },
 ]
 
+[[package]]
+name = "starlette"
+version = "0.46.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "anyio" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/04/1b/52b27f2e13ceedc79a908e29eac426a63465a1a01248e5f24aa36a62aeb3/starlette-0.46.1.tar.gz", hash = "sha256:3c88d58ee4bd1bb807c0d1acb381838afc7752f9ddaec81bbe4383611d833230", size = 2580102 }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/a0/4b/528ccf7a982216885a1ff4908e886b8fb5f19862d1962f56a3fce2435a70/starlette-0.46.1-py3-none-any.whl", hash = "sha256:77c74ed9d2720138b25875133f3a2dae6d854af2ec37dceb56aef370c1d8a227", size = 71995 },
+]
+
 [[package]]
 name = "sympy"
 version = "1.13.1"
@@ -1081,6 +1153,28 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/88/8b/d60c0491ab63634763be1537ad488694d316ddc4a20eaadd639cedc53971/torch-2.6.0-cp313-none-macosx_11_0_arm64.whl", hash = "sha256:ff96f4038f8af9f7ec4231710ed4549da1bdebad95923953a25045dcf6fd87e2", size = 66536783 },
 ]
 
+[[package]]
+name = "torchaudio"
+version = "2.6.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "torch" },
+]
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/a9/30/bba293c8300245a09b7f82d3cfc04aee1950228da49c6cdd637d1145b6f5/torchaudio-2.6.0-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:c12fc41241b8dfce3ccc1917f1c81a0f92f532d9917706600046f1eb21d2d765", size = 1815253 },
+    { url = "https://files.pythonhosted.org/packages/3e/00/2c69d436c613043f3051210d2f84a4c9062a815fa609c5f54d25ea8bfd07/torchaudio-2.6.0-cp311-cp311-manylinux1_x86_64.whl", hash = "sha256:377b177a3d683a9163e4cab5a06f0346dac9ff96fa527477338fd90fc6a2a4b6", size = 3382518 },
+    { url = "https://files.pythonhosted.org/packages/f5/b8/7d4dbbf6b505caddbfccd38e2882e47a791310b32b347f977a0a66efbf80/torchaudio-2.6.0-cp311-cp311-manylinux_2_28_aarch64.whl", hash = "sha256:0f0db5c997d031c34066d8be1c0ce7d2a1f2b6c016a92885b20b00bfeb17b753", size = 1652980 },
+    { url = "https://files.pythonhosted.org/packages/1f/31/417d6955585be76842e9b0159d3801c0b5f9a4ea0db39db1a72bc262c861/torchaudio-2.6.0-cp311-cp311-win_amd64.whl", hash = "sha256:52182f6de4e7b342d139e54b703185d428de9cce3c4cf914a9b2ab2359d192a3", size = 2454430 },
+    { url = "https://files.pythonhosted.org/packages/ac/4a/d71b932bda4171970bdf4997541b5c778daa0e2967ed5009d207fca86ded/torchaudio-2.6.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:7d0e4b08c42325bf4b887de9a25c44ed882997001740e1bd7d901f65581cf1ab", size = 1812899 },
+    { url = "https://files.pythonhosted.org/packages/ed/aa/9082e715a673dd8e22b6a60cec7f301e897406023672b2090f8bcd8a5959/torchaudio-2.6.0-cp312-cp312-manylinux1_x86_64.whl", hash = "sha256:715aa21f6bdbd085454c313ae3a2c7cc07bf2e8cf05752f819afb5b4c57f4e6f", size = 3379510 },
+    { url = "https://files.pythonhosted.org/packages/f2/e7/0bcb2e33f4bdec69477344eccfe25c515b90496888095e99f837ea422089/torchaudio-2.6.0-cp312-cp312-manylinux_2_28_aarch64.whl", hash = "sha256:6291d9507dc1d6b4ffe8843fbfb201e6c8270dd8c42ad70bb76226c0ebdcad56", size = 1653523 },
+    { url = "https://files.pythonhosted.org/packages/80/95/29e917905328337c7b104ce81f3bb5e2ad8dc70af2edf1d43f67eb621513/torchaudio-2.6.0-cp312-cp312-win_amd64.whl", hash = "sha256:86d6239792bf94741a41acd6fe3d549faaf0d50e7275d17d076a190bd007e2f9", size = 2449191 },
+    { url = "https://files.pythonhosted.org/packages/fb/73/861afa5864e95fbf42b693e0359b2bf0177b6b5f4274fa4472fd51e5298e/torchaudio-2.6.0-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:66f2e0bd5ab56fd81419d2f5afb74a9a70141688594646441756c8c24f424a73", size = 1813188 },
+    { url = "https://files.pythonhosted.org/packages/d2/f0/daffd9afa60bd835a2d7980eddfe44524adcb3ee0837486ceae4cd1f68e2/torchaudio-2.6.0-cp313-cp313-manylinux1_x86_64.whl", hash = "sha256:52f15185349c370fc1faa84e8b8b2782c007472db9d586a16bba314130b322f2", size = 3380706 },
+    { url = "https://files.pythonhosted.org/packages/94/7b/887b91372e34119aa140cf67614e5ba901bf6a0db86f2c39e30ff71eec54/torchaudio-2.6.0-cp313-cp313-manylinux_2_28_aarch64.whl", hash = "sha256:b521ea9618fb4c29a6f8071628170c222291f46a48a3bf424cfeb488f54af714", size = 1653553 },
+    { url = "https://files.pythonhosted.org/packages/55/c8/3010878a5e7f15d89450e22769697173c6dc244a0647ddc5386c28b6dacc/torchaudio-2.6.0-cp313-cp313-win_amd64.whl", hash = "sha256:393fa74ec40d167f0170728ea21c9b5e0f830648fd02df7db2bf7e62f64245ec", size = 2449350 },
+]
+
 [[package]]
 name = "tqdm"
 version = "4.67.1"
@@ -1099,12 +1193,15 @@ version = "0.1.0"
 source = { virtual = "." }
 dependencies = [
     { name = "av" },
+    { name = "fastapi" },
     { name = "librosa" },
     { name = "numpy" },
     { name = "onnxruntime" },
     { name = "pyaudio" },
     { name = "pydantic" },
+    { name = "pydub" },
     { name = "setuptools" },
+    { name = "silero-vad" },
     { name = "soundfile" },
     { name = "torch" },
     { name = "tqdm" },
@@ -1115,12 +1212,15 @@ dependencies = [
 [package.metadata]
 requires-dist = [
     { name = "av", specifier = ">=14.2.0" },
+    { name = "fastapi", specifier = ">=0.115.12" },
     { name = "librosa", specifier = ">=0.11.0" },
     { name = "numpy", specifier = ">=2.1.3" },
     { name = "onnxruntime", specifier = ">=1.21.0" },
     { name = "pyaudio", specifier = ">=0.2.14" },
     { name = "pydantic", specifier = ">=2.11.2" },
+    { name = "pydub", specifier = ">=0.25.1" },
     { name = "setuptools", specifier = ">=78.1.0" },
+    { name = "silero-vad", specifier = ">=5.1.2" },
     { name = "soundfile", specifier = ">=0.13.1" },
     { name = "torch", specifier = ">=2.6.0" },
     { name = "tqdm", specifier = ">=4.67.1" },