Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,4 +1,5 @@
|
|
| 1 |
import os
|
|
|
|
| 2 |
import spaces
|
| 3 |
# only debug for hf now
|
| 4 |
REPO_TYPE = "hf"
|
|
@@ -376,18 +377,62 @@ def trim_audio(audio_path, start_time, end_time):
|
|
| 376 |
raise gr.Error(f"Error trimming audio: {str(e)}")
|
| 377 |
|
| 378 |
def save_transcription(transcription):
|
| 379 |
-
"""
|
| 380 |
-
|
| 381 |
-
|
| 382 |
-
Args:
|
| 383 |
-
transcription (str): The transcription text.
|
| 384 |
-
|
| 385 |
-
Returns:
|
| 386 |
-
str: The path to the transcription file.
|
| 387 |
-
"""
|
| 388 |
-
with tempfile.NamedTemporaryFile(delete=False, suffix='.txt', mode='w', encoding='utf-8') as temp_file:
|
| 389 |
temp_file.write(transcription)
|
| 390 |
-
logging.info(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 391 |
return temp_file.name
|
| 392 |
|
| 393 |
def get_model_options(pipeline_type):
|
|
@@ -445,7 +490,7 @@ def transcribe_audio(audio_input, audio_url, proxy_url, proxy_username, proxy_pa
|
|
| 445 |
verbose_messages = f"Starting transcription with parameters:\nPipeline Type: {pipeline_type}\nModel ID: {model_id}\nDownload Method: {download_method}\n"
|
| 446 |
|
| 447 |
if verbose:
|
| 448 |
-
yield verbose_messages, "", None
|
| 449 |
|
| 450 |
# Determine the audio source
|
| 451 |
audio_path = None
|
|
@@ -461,16 +506,16 @@ def transcribe_audio(audio_input, audio_url, proxy_url, proxy_username, proxy_pa
|
|
| 461 |
if not audio_path:
|
| 462 |
error_msg = f"Error downloading audio from {audio_url} using method {download_method}. Check logs for details."
|
| 463 |
logging.error(error_msg)
|
| 464 |
-
yield verbose_messages + error_msg, "", None
|
| 465 |
return
|
| 466 |
else:
|
| 467 |
verbose_messages += f"Successfully downloaded audio from {audio_url}\n"
|
| 468 |
if verbose:
|
| 469 |
-
yield verbose_messages, "", None
|
| 470 |
else:
|
| 471 |
error_msg = "No audio source provided. Please upload an audio file, record audio, or enter a URL."
|
| 472 |
logging.error(error_msg)
|
| 473 |
-
yield verbose_messages + error_msg, "", None
|
| 474 |
return
|
| 475 |
|
| 476 |
# Convert start_time and end_time to float or None
|
|
@@ -482,7 +527,7 @@ def transcribe_audio(audio_input, audio_url, proxy_url, proxy_username, proxy_pa
|
|
| 482 |
is_temp_file = True # The trimmed audio is a temporary file
|
| 483 |
verbose_messages += f"Audio trimmed from {start_time} to {end_time}\n"
|
| 484 |
if verbose:
|
| 485 |
-
yield verbose_messages, "", None
|
| 486 |
|
| 487 |
# Model caching
|
| 488 |
model_key = (pipeline_type, model_id)
|
|
@@ -514,7 +559,7 @@ def transcribe_audio(audio_input, audio_url, proxy_url, proxy_username, proxy_pa
|
|
| 514 |
else:
|
| 515 |
error_msg = "Invalid pipeline type. Only 'sensevoice' is supported."
|
| 516 |
logging.error(error_msg)
|
| 517 |
-
yield verbose_messages + error_msg, "", None
|
| 518 |
return
|
| 519 |
loaded_models[model_key] = model
|
| 520 |
|
|
@@ -552,22 +597,31 @@ def transcribe_audio(audio_input, audio_url, proxy_url, proxy_username, proxy_pa
|
|
| 552 |
transcription_time = end_time_perf - start_time_perf
|
| 553 |
audio_file_size = os.path.getsize(audio_path) / (1024 * 1024)
|
| 554 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 555 |
metrics_output = (
|
| 556 |
f"Transcription time: {transcription_time:.2f} seconds\n"
|
| 557 |
f"Audio file size: {audio_file_size:.2f} MB\n"
|
|
|
|
|
|
|
| 558 |
)
|
| 559 |
|
| 560 |
-
# Save
|
| 561 |
transcription_file = save_transcription(transcription)
|
|
|
|
| 562 |
|
| 563 |
# Always yield the final result, regardless of verbose setting
|
| 564 |
final_metrics = verbose_messages + metrics_output
|
| 565 |
-
yield final_metrics, transcription, transcription_file
|
| 566 |
|
| 567 |
except Exception as e:
|
| 568 |
error_msg = f"An error occurred during transcription: {str(e)}"
|
| 569 |
logging.error(error_msg)
|
| 570 |
-
|
|
|
|
| 571 |
|
| 572 |
finally:
|
| 573 |
# Clean up temporary audio files
|
|
@@ -618,7 +672,8 @@ with gr.Blocks() as iface:
|
|
| 618 |
with gr.Row():
|
| 619 |
metrics_output = gr.Textbox(label="Transcription Metrics and Verbose Messages", lines=10)
|
| 620 |
transcription_output = gr.Textbox(label="Transcription", lines=10)
|
| 621 |
-
transcription_file = gr.File(label="Download
|
|
|
|
| 622 |
|
| 623 |
def update_model_dropdown(pipeline_type):
|
| 624 |
"""
|
|
@@ -652,7 +707,7 @@ with gr.Blocks() as iface:
|
|
| 652 |
transcribe_button.click(
|
| 653 |
transcribe_with_progress,
|
| 654 |
inputs=[audio_input, audio_url, proxy_url, proxy_username, proxy_password, pipeline_type, model_id, download_method, start_time, end_time, verbose],
|
| 655 |
-
outputs=[metrics_output, transcription_output, transcription_file]
|
| 656 |
)
|
| 657 |
|
| 658 |
# Note: For examples, users should use local audio files or upload their own files
|
|
|
|
| 1 |
import os
|
| 2 |
+
import re
|
| 3 |
import spaces
|
| 4 |
# only debug for hf now
|
| 5 |
REPO_TYPE = "hf"
|
|
|
|
| 377 |
raise gr.Error(f"Error trimming audio: {str(e)}")
|
| 378 |
|
| 379 |
def save_transcription(transcription):
|
| 380 |
+
"""Save transcription text to a temporary TXT file."""
|
| 381 |
+
with tempfile.NamedTemporaryFile(delete=False, suffix=".txt", mode="w", encoding="utf-8") as temp_file:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 382 |
temp_file.write(transcription)
|
| 383 |
+
logging.info("Transcription saved to: %s", temp_file.name)
|
| 384 |
+
return temp_file.name
|
| 385 |
+
|
| 386 |
+
|
| 387 |
+
def _format_srt_time(seconds):
|
| 388 |
+
"""Format seconds as SRT timestamp: HH:MM:SS,mmm."""
|
| 389 |
+
total_ms = max(0, int(round(float(seconds) * 1000)))
|
| 390 |
+
hours, rem = divmod(total_ms, 3_600_000)
|
| 391 |
+
minutes, rem = divmod(rem, 60_000)
|
| 392 |
+
secs, millis = divmod(rem, 1000)
|
| 393 |
+
return f"{hours:02d}:{minutes:02d}:{secs:02d},{millis:03d}"
|
| 394 |
+
|
| 395 |
+
|
| 396 |
+
def _split_subtitle_text(text, max_chars=42):
|
| 397 |
+
"""Split transcript into subtitle-sized chunks."""
|
| 398 |
+
text = (text or "").strip()
|
| 399 |
+
if not text:
|
| 400 |
+
return []
|
| 401 |
+
sentences = re.split(r"(?<=[。!?!?;;])\s*|\n+", text)
|
| 402 |
+
chunks = []
|
| 403 |
+
for sentence in sentences:
|
| 404 |
+
sentence = sentence.strip()
|
| 405 |
+
while len(sentence) > max_chars:
|
| 406 |
+
cut = max_chars
|
| 407 |
+
for pos in range(max_chars, max(8, max_chars - 12), -1):
|
| 408 |
+
if sentence[pos - 1:pos] in ",,、 ":
|
| 409 |
+
cut = pos
|
| 410 |
+
break
|
| 411 |
+
chunks.append(sentence[:cut].strip())
|
| 412 |
+
sentence = sentence[cut:].strip()
|
| 413 |
+
if sentence:
|
| 414 |
+
chunks.append(sentence)
|
| 415 |
+
return chunks
|
| 416 |
+
|
| 417 |
+
|
| 418 |
+
def save_srt(transcription, duration_seconds):
|
| 419 |
+
"""Save an SRT file. Timings are estimated when the ASR model provides no timestamps."""
|
| 420 |
+
chunks = _split_subtitle_text(transcription) or [""]
|
| 421 |
+
duration = max(float(duration_seconds or 0), 1.0)
|
| 422 |
+
weights = [max(len(chunk), 1) for chunk in chunks]
|
| 423 |
+
total_weight = sum(weights)
|
| 424 |
+
cursor = 0.0
|
| 425 |
+
blocks = []
|
| 426 |
+
for index, (chunk, weight) in enumerate(zip(chunks, weights), start=1):
|
| 427 |
+
start_time = cursor
|
| 428 |
+
end_time = duration if index == len(chunks) else duration * sum(weights[:index]) / total_weight
|
| 429 |
+
if end_time <= start_time:
|
| 430 |
+
end_time = start_time + 0.5
|
| 431 |
+
blocks.append(f"{index}\n{_format_srt_time(start_time)} --> {_format_srt_time(end_time)}\n{chunk}")
|
| 432 |
+
cursor = end_time
|
| 433 |
+
with tempfile.NamedTemporaryFile(delete=False, suffix=".srt", mode="w", encoding="utf-8-sig", newline="\n") as temp_file:
|
| 434 |
+
temp_file.write("\n\n".join(blocks) + "\n")
|
| 435 |
+
logging.info("SRT subtitles saved to: %s", temp_file.name)
|
| 436 |
return temp_file.name
|
| 437 |
|
| 438 |
def get_model_options(pipeline_type):
|
|
|
|
| 490 |
verbose_messages = f"Starting transcription with parameters:\nPipeline Type: {pipeline_type}\nModel ID: {model_id}\nDownload Method: {download_method}\n"
|
| 491 |
|
| 492 |
if verbose:
|
| 493 |
+
yield verbose_messages, "", None, None
|
| 494 |
|
| 495 |
# Determine the audio source
|
| 496 |
audio_path = None
|
|
|
|
| 506 |
if not audio_path:
|
| 507 |
error_msg = f"Error downloading audio from {audio_url} using method {download_method}. Check logs for details."
|
| 508 |
logging.error(error_msg)
|
| 509 |
+
yield verbose_messages + error_msg, "", None, None
|
| 510 |
return
|
| 511 |
else:
|
| 512 |
verbose_messages += f"Successfully downloaded audio from {audio_url}\n"
|
| 513 |
if verbose:
|
| 514 |
+
yield verbose_messages, "", None, None
|
| 515 |
else:
|
| 516 |
error_msg = "No audio source provided. Please upload an audio file, record audio, or enter a URL."
|
| 517 |
logging.error(error_msg)
|
| 518 |
+
yield verbose_messages + error_msg, "", None, None
|
| 519 |
return
|
| 520 |
|
| 521 |
# Convert start_time and end_time to float or None
|
|
|
|
| 527 |
is_temp_file = True # The trimmed audio is a temporary file
|
| 528 |
verbose_messages += f"Audio trimmed from {start_time} to {end_time}\n"
|
| 529 |
if verbose:
|
| 530 |
+
yield verbose_messages, "", None, None
|
| 531 |
|
| 532 |
# Model caching
|
| 533 |
model_key = (pipeline_type, model_id)
|
|
|
|
| 559 |
else:
|
| 560 |
error_msg = "Invalid pipeline type. Only 'sensevoice' is supported."
|
| 561 |
logging.error(error_msg)
|
| 562 |
+
yield verbose_messages + error_msg, "", None, None
|
| 563 |
return
|
| 564 |
loaded_models[model_key] = model
|
| 565 |
|
|
|
|
| 597 |
transcription_time = end_time_perf - start_time_perf
|
| 598 |
audio_file_size = os.path.getsize(audio_path) / (1024 * 1024)
|
| 599 |
|
| 600 |
+
try:
|
| 601 |
+
audio_duration = len(AudioSegment.from_file(audio_path)) / 1000.0
|
| 602 |
+
except Exception:
|
| 603 |
+
audio_duration = max(float(end_time or 0), 1.0)
|
| 604 |
+
|
| 605 |
metrics_output = (
|
| 606 |
f"Transcription time: {transcription_time:.2f} seconds\n"
|
| 607 |
f"Audio file size: {audio_file_size:.2f} MB\n"
|
| 608 |
+
f"Audio duration: {audio_duration:.2f} seconds\n"
|
| 609 |
+
"SRT timing is estimated from text/audio duration when the model provides no timestamps.\n"
|
| 610 |
)
|
| 611 |
|
| 612 |
+
# Save both plain-text transcript and SRT subtitles.
|
| 613 |
transcription_file = save_transcription(transcription)
|
| 614 |
+
srt_file = save_srt(transcription, audio_duration)
|
| 615 |
|
| 616 |
# Always yield the final result, regardless of verbose setting
|
| 617 |
final_metrics = verbose_messages + metrics_output
|
| 618 |
+
yield final_metrics, transcription, transcription_file, srt_file
|
| 619 |
|
| 620 |
except Exception as e:
|
| 621 |
error_msg = f"An error occurred during transcription: {str(e)}"
|
| 622 |
logging.error(error_msg)
|
| 623 |
+
prefix = locals().get("verbose_messages", "")
|
| 624 |
+
yield prefix + error_msg, "", None, None
|
| 625 |
|
| 626 |
finally:
|
| 627 |
# Clean up temporary audio files
|
|
|
|
| 672 |
with gr.Row():
|
| 673 |
metrics_output = gr.Textbox(label="Transcription Metrics and Verbose Messages", lines=10)
|
| 674 |
transcription_output = gr.Textbox(label="Transcription", lines=10)
|
| 675 |
+
transcription_file = gr.File(label="Download TXT")
|
| 676 |
+
srt_file_output = gr.File(label="Download SRT Subtitles")
|
| 677 |
|
| 678 |
def update_model_dropdown(pipeline_type):
|
| 679 |
"""
|
|
|
|
| 707 |
transcribe_button.click(
|
| 708 |
transcribe_with_progress,
|
| 709 |
inputs=[audio_input, audio_url, proxy_url, proxy_username, proxy_password, pipeline_type, model_id, download_method, start_time, end_time, verbose],
|
| 710 |
+
outputs=[metrics_output, transcription_output, transcription_file, srt_file_output]
|
| 711 |
)
|
| 712 |
|
| 713 |
# Note: For examples, users should use local audio files or upload their own files
|