Heidel Medina.
commited on
Commit
·
e935b66
1
Parent(s):
3d86161
Added advanced settings feature to main.py
Browse files
main.py
CHANGED
@@ -9,35 +9,89 @@ import tempfile
|
|
9 |
import re
|
10 |
import textwrap
|
11 |
|
12 |
-
def process_media(
|
|
|
|
|
|
|
|
|
13 |
if upload is None:
|
14 |
return None, None, None, None, "No file uploaded."
|
15 |
|
16 |
temp_path = upload.name
|
|
|
|
|
17 |
|
18 |
-
if
|
19 |
-
|
|
|
20 |
else:
|
21 |
-
|
22 |
-
|
23 |
-
|
24 |
-
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
-
|
33 |
-
|
34 |
-
|
35 |
-
|
36 |
-
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
41 |
transcript_txt = result.to_txt()
|
42 |
|
43 |
mime, _ = mimetypes.guess_type(temp_path)
|
@@ -45,7 +99,59 @@ def process_media(model_size, source_lang, upload, model_type):
|
|
45 |
video_out = temp_path if mime and mime.startswith("video") else None
|
46 |
|
47 |
return audio_out, video_out, transcript_txt, srt_file_path, None
|
48 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
49 |
WHISPER_LANGUAGES = [
|
50 |
("Afrikaans", "af"),
|
51 |
("Albanian", "sq"),
|
@@ -290,7 +396,11 @@ with gr.Blocks() as interface:
|
|
290 |
|
291 |
submit_btn.click(
|
292 |
fn=process_media,
|
293 |
-
inputs=[
|
|
|
|
|
|
|
|
|
294 |
outputs=[audio_output, video_output, transcript_output, srt_output]
|
295 |
)
|
296 |
|
|
|
9 |
import re
|
10 |
import textwrap
|
11 |
|
12 |
+
def process_media(
|
13 |
+
model_size, source_lang, upload, model_type,
|
14 |
+
max_chars, max_words, extend_in, extend_out, collapse_gaps,
|
15 |
+
max_lines_per_segment, line_penalty, longest_line_char_penalty, *args
|
16 |
+
):
|
17 |
if upload is None:
|
18 |
return None, None, None, None, "No file uploaded."
|
19 |
|
20 |
temp_path = upload.name
|
21 |
+
base_path = os.path.splitext(temp_path)[0]
|
22 |
+
word_transcription_path = base_path + '.json'
|
23 |
|
24 |
+
if os.path.exists(word_transcription_path):
|
25 |
+
print(f"Transcription data file found at {word_transcription_path}")
|
26 |
+
result = stable_whisper.WhisperResult(word_transcription_path)
|
27 |
else:
|
28 |
+
print(f"Can't find transcription data file at {word_transcription_path}. Starting transcribing ...")
|
29 |
+
if model_type == "faster whisper":
|
30 |
+
model = stable_whisper.load_faster_whisper(model_size, device="cuda")
|
31 |
+
else:
|
32 |
+
model = stable_whisper.load_model(model_size, device="cuda")
|
33 |
+
try:
|
34 |
+
result = model.transcribe(temp_path, language=source_lang, vad=True, regroup=False, denoiser="demucs")
|
35 |
+
except Exception as e:
|
36 |
+
return None, None, None, None, f"Transcription failed: {e}"
|
37 |
+
result.save_as_json(word_transcription_path)
|
38 |
+
|
39 |
+
if max_chars or max_words:
|
40 |
+
result.split_by_length(
|
41 |
+
max_chars=int(max_chars) if max_chars else None,
|
42 |
+
max_words=int(max_words) if max_words else None
|
43 |
+
)
|
44 |
+
|
45 |
+
# ----- Perform segment time extensions and anti-flickering (=closing the gaps) -----
|
46 |
+
extend_start = float(extend_in) if extend_in else 0.0
|
47 |
+
extend_end = float(extend_out) if extend_out else 0.0
|
48 |
+
collapse_gaps_under = float(collapse_gaps) if collapse_gaps else 0.0
|
49 |
+
|
50 |
+
for i in range(len(result) - 1):
|
51 |
+
cur = result[i]
|
52 |
+
next = result[i+1]
|
53 |
+
|
54 |
+
if next.start - cur.end < extend_start + extend_end:
|
55 |
+
# Not enough time to add the entire desired extensions -> add proportionally
|
56 |
+
k = extend_end / (extend_start + extend_end) if (extend_start + extend_end) > 0 else 0
|
57 |
+
mid = cur.end * (1 - k) + next.start * k
|
58 |
+
cur.end = next.start = mid
|
59 |
+
else:
|
60 |
+
# Add full desired extensions
|
61 |
+
cur.end += extend_end
|
62 |
+
next.start -= extend_start
|
63 |
+
|
64 |
+
if next.start - cur.end <= collapse_gaps_under:
|
65 |
+
cur.end = next.start = (cur.end + next.start) / 2
|
66 |
+
|
67 |
+
if result:
|
68 |
+
result[0].start = max(0, result[0].start - extend_start)
|
69 |
+
result[-1].end += extend_end
|
70 |
+
|
71 |
+
#for seg in result:
|
72 |
+
# seg.text = optimize_text(
|
73 |
+
# seg.text,
|
74 |
+
# int(max_lines_per_segment) if max_lines_per_segment else 3,
|
75 |
+
# float(line_penalty) if line_penalty else 22.01,
|
76 |
+
# float(longest_line_char_penalty) if longest_line_char_penalty else 1.0
|
77 |
+
# )
|
78 |
+
|
79 |
+
# Use custom SRT block output
|
80 |
+
subtitles_path = tempfile.NamedTemporaryFile(delete=False, suffix=".srt", mode="w", encoding="utf-8").name
|
81 |
+
result_to_any(
|
82 |
+
result=result,
|
83 |
+
filepath=subtitles_path,
|
84 |
+
filetype='srt',
|
85 |
+
segments2blocks=lambda segments: segments2blocks(
|
86 |
+
segments,
|
87 |
+
int(max_lines_per_segment) if max_lines_per_segment else 3,
|
88 |
+
float(line_penalty) if line_penalty else 22.01,
|
89 |
+
float(longest_line_char_penalty) if longest_line_char_penalty else 1.0
|
90 |
+
),
|
91 |
+
word_level=False,
|
92 |
+
)
|
93 |
+
srt_file_path = subtitles_path
|
94 |
+
|
95 |
transcript_txt = result.to_txt()
|
96 |
|
97 |
mime, _ = mimetypes.guess_type(temp_path)
|
|
|
99 |
video_out = temp_path if mime and mime.startswith("video") else None
|
100 |
|
101 |
return audio_out, video_out, transcript_txt, srt_file_path, None
|
102 |
+
|
103 |
+
def optimize_text(text, max_lines_per_segment, line_penalty, longest_line_char_penalty):
|
104 |
+
text = text.strip()
|
105 |
+
words = text.split()
|
106 |
+
|
107 |
+
# Compute prefix sums
|
108 |
+
psum = [0]
|
109 |
+
for w in words:
|
110 |
+
psum += [psum[-1] + len(w) + 1] # +1 because of spaces
|
111 |
+
|
112 |
+
bestScore = 10 ** 30
|
113 |
+
bestSplit = None
|
114 |
+
|
115 |
+
def backtrack(level, wordsUsed, maxLineLength, split):
|
116 |
+
nonlocal bestScore, bestSplit
|
117 |
+
|
118 |
+
if wordsUsed == len(words):
|
119 |
+
score = level * line_penalty + maxLineLength * longest_line_char_penalty
|
120 |
+
if score < bestScore:
|
121 |
+
bestScore = score
|
122 |
+
bestSplit = split
|
123 |
+
return
|
124 |
+
|
125 |
+
if level + 1 == max_lines_per_segment:
|
126 |
+
backtrack(
|
127 |
+
level + 1, len(words),
|
128 |
+
max(maxLineLength, psum[len(words)] - psum[wordsUsed] - 1),
|
129 |
+
split + [words[wordsUsed:]]
|
130 |
+
)
|
131 |
+
return
|
132 |
+
|
133 |
+
for levelWords in range(1, len(words) - wordsUsed + 1):
|
134 |
+
backtrack(
|
135 |
+
level + 1, wordsUsed + levelWords,
|
136 |
+
max(maxLineLength, psum[wordsUsed + levelWords] - psum[wordsUsed] - 1),
|
137 |
+
split + [words[wordsUsed:wordsUsed + levelWords]]
|
138 |
+
)
|
139 |
+
|
140 |
+
backtrack(0, 0, 0, [])
|
141 |
+
|
142 |
+
optimized = '\n'.join(' '.join(words) for words in bestSplit)
|
143 |
+
return optimized
|
144 |
+
|
145 |
+
def segment2optimizedsrtblock(segment: dict, idx: int, max_lines_per_segment, line_penalty, longest_line_char_penalty, strip=True) -> str:
|
146 |
+
return f'{idx}\n{sec2srt(segment["start"])} --> {sec2srt(segment["end"])}\n' \
|
147 |
+
f'{optimize_text(segment["text"], max_lines_per_segment, line_penalty, longest_line_char_penalty)}'
|
148 |
+
|
149 |
+
def segments2blocks(segments, max_lines_per_segment, line_penalty, longest_line_char_penalty):
|
150 |
+
return '\n\n'.join(
|
151 |
+
segment2optimizedsrtblock(s, i, max_lines_per_segment, line_penalty, longest_line_char_penalty, strip=True)
|
152 |
+
for i, s in enumerate(segments)
|
153 |
+
)
|
154 |
+
|
155 |
WHISPER_LANGUAGES = [
|
156 |
("Afrikaans", "af"),
|
157 |
("Albanian", "sq"),
|
|
|
396 |
|
397 |
submit_btn.click(
|
398 |
fn=process_media,
|
399 |
+
inputs=[
|
400 |
+
model_size, source_lang, file_input, model_type,
|
401 |
+
max_chars, max_words, extend_in, extend_out, collapse_gaps,
|
402 |
+
max_lines_per_segment, line_penalty, longest_line_char_penalty
|
403 |
+
],
|
404 |
outputs=[audio_output, video_output, transcript_output, srt_output]
|
405 |
)
|
406 |
|