Upload app.py

app.py

@@ -128,27 +128,42 @@ def ms_to_srt_time(ms):
     return f"{hours:02d}:{minutes:02d}:{seconds:02d},{milliseconds:03d}"
 
 
-def generate_srt(sentence_info, clean=True):
+def generate_srt(sentence_info):
     if not sentence_info:
         return ""
     srt_lines = []
     for i, seg in enumerate(sentence_info, 1):
         start = ms_to_srt_time(seg.get("start", 0))
         end = ms_to_srt_time(seg.get("end", 0))
-        text = seg.get("text", "")
-        if clean:
-            text = re.sub(r"<\|[^>]*\|?>", "", text)
-        text = text.strip()
+        text = re.sub(r"<\|[^>]*\|?>", "", seg.get("text", "")).strip()
         if text:
             srt_lines.append(f"{i}\n{start} --> {end}\n{text}\n")
     return "\n".join(srt_lines).strip()
 
 
+def apply_format(raw_text, sentence_info, output_format):
+    if output_format == "纯净文本":
+        return re.sub(r"<\|[^>]*\|?>", "", raw_text).strip()
+    elif output_format == "原始富文本":
+        return raw_text
+    elif output_format == "Emoji 格式":
+        return format_to_emoji(raw_text)
+    elif output_format == "SRT 字幕":
+        srt_result = generate_srt(sentence_info)
+        return (
+            srt_result if srt_result else re.sub(r"<\|[^>]*\|?>", "", raw_text).strip()
+        )
+    elif output_format == "ALL_IN_ONE":
+        srt_text = generate_srt(sentence_info)
+        return f"{raw_text}\n===SRT_DELIMITER===\n{srt_text}"
+    return raw_text
+
+
 def model_inference(
     audio_input, language, output_format, use_itn, merge_vad, merge_length, ban_emo_unk
 ):
     if audio_input is None:
-        return "错误：请上传或录制音频。"
+        return "错误：请上传或录制音频。", None
     fs, input_wav = audio_input
     if input_wav.dtype in [np.int16, np.int32]:
         input_wav = input_wav.astype(np.float32) / np.iinfo(input_wav.dtype).max
@@ -159,6 +174,7 @@ def model_inference(
     if fs != 16000:
         resampler = torchaudio.transforms.Resample(orig_freq=fs, new_freq=16000)
         input_wav = resampler(torch.from_numpy(input_wav).to(torch.float32)).numpy()
+
     res = model.generate(
         input=input_wav,
         cache={},
@@ -172,25 +188,21 @@ def model_inference(
     )
     raw_text = res[0].get("text", "")
     if not raw_text:
-        return "未能识别出文本。"
+        return "未能识别出文本。", None
 
     sentence_info = res[0].get("sentence_info", [])
+    cache_state = {"raw_text": raw_text, "sentence_info": sentence_info}
 
-    if output_format == "纯净文本":
-        return re.sub(r"<\|[^>]*\|?>", "", raw_text).strip()
-    elif output_format == "原始富文本":
-        return raw_text
-    elif output_format == "Emoji 格式":
-        return format_to_emoji(raw_text)
-    elif output_format == "SRT 字幕":
-        srt_result = generate_srt(sentence_info, clean=True)
-        return (
-            srt_result if srt_result else re.sub(r"<\|[^>]*\|?>", "", raw_text).strip()
-        )
-    elif output_format == "MASTER_DATA":
-        srt_raw = generate_srt(sentence_info, clean=False)
-        return srt_raw if srt_raw else raw_text
-    return raw_text
+    result_text = apply_format(raw_text, sentence_info, output_format)
+    return result_text, cache_state
+
+
+def on_format_change(output_format, cache_state):
+    if not cache_state or "raw_text" not in cache_state:
+        return gr.update()
+    return apply_format(
+        cache_state["raw_text"], cache_state.get("sentence_info", []), output_format
+    )
 
 
 html_intro = """<div style="text-align: center; font-family: var(--font-sans);"><h1 style="font-size: 28px;">SenseVoice-Small 语音识别模型</h1><p>SenseVoice 是具有音频理解能力的音频基础模型，包括语音识别（ASR）、语种识别（LID）、语音情感识别（SER）和声学事件分类（AEC）或声学事件检测（AED）。</p><p style="font-size: small; color: #888;">支持 MP3, WAV, FLAC, M4A 等常见音频格式。</p></div>"""
@@ -231,6 +243,8 @@ lang_options = [
 
 with gr.Blocks(theme=gr.themes.Soft(), css=custom_css) as demo:
     gr.HTML(html_intro)
+    cached_data = gr.State(value=None)
+
     with gr.Row():
         with gr.Column(scale=2):
             audio_inputs = gr.Audio(label="上传/录制音频")
@@ -258,6 +272,8 @@ with gr.Blocks(theme=gr.themes.Soft(), css=custom_css) as demo:
             fn_button = gr.Button("开始识别", variant="primary")
         with gr.Column(scale=3):
             text_outputs = gr.Textbox(label="识别结果", lines=25, show_copy_button=True)
+
+    # 点击识别跑 GPU 模型，并将识别结果缓存到 cached_data
     fn_button.click(
         fn=model_inference,
         inputs=[
@@ -269,8 +285,15 @@ with gr.Blocks(theme=gr.themes.Soft(), css=custom_css) as demo:
             merge_length_slider,
             ban_emo_unk_checkbox,
         ],
-        outputs=text_outputs,
+        outputs=[text_outputs, cached_data],
         api_name="model_inference",
     )
 
+    # 识别完成后直接切下拉框，0.001 秒即时切换格式（不重新跑 GPU）
+    output_format_dropdown.change(
+        fn=on_format_change,
+        inputs=[output_format_dropdown, cached_data],
+        outputs=text_outputs,
+    )
+
 demo.launch(server_name="0.0.0.0", server_port=7860)