Ma/auto-caption
1
0
Fork 0
mirror of https://github.com/HiMeditator/auto-caption.git synced 2026-02-17 14:14:53 +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

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()