Spaces:
Runtime error
Runtime error
Matthijs Hollemans
commited on
Commit
•
b1828a3
1
Parent(s):
dbc8f56
let's go!
Browse files- .gitattributes +3 -0
- README.md +2 -2
- app.py +149 -0
- background.png +0 -0
- requirements.txt +8 -0
.gitattributes
CHANGED
@@ -32,3 +32,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
32 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
33 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
34 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
32 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
33 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
34 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
35 |
+
*.wav filter=lfs diff=lfs merge=lfs -text
|
36 |
+
*.mp3 filter=lfs diff=lfs merge=lfs -text
|
37 |
+
*.ttf filter=lfs diff=lfs merge=lfs -text
|
README.md
CHANGED
@@ -1,6 +1,6 @@
|
|
1 |
---
|
2 |
-
title: Whisper Word Timestamps
|
3 |
-
emoji:
|
4 |
colorFrom: yellow
|
5 |
colorTo: indigo
|
6 |
sdk: gradio
|
|
|
1 |
---
|
2 |
+
title: Whisper Word-Level Timestamps
|
3 |
+
emoji: 💭⏰
|
4 |
colorFrom: yellow
|
5 |
colorTo: indigo
|
6 |
sdk: gradio
|
app.py
ADDED
@@ -0,0 +1,149 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
import librosa
|
3 |
+
import numpy as np
|
4 |
+
import moviepy.editor as mpy
|
5 |
+
|
6 |
+
from PIL import Image, ImageDraw, ImageFont
|
7 |
+
from transformers import pipeline
|
8 |
+
|
9 |
+
|
10 |
+
fps = 25
|
11 |
+
max_duration = 60 # seconds
|
12 |
+
video_width = 640
|
13 |
+
video_height = 480
|
14 |
+
margin_left = 20
|
15 |
+
margin_right = 20
|
16 |
+
margin_top = 20
|
17 |
+
line_height = 44
|
18 |
+
|
19 |
+
background_image = Image.open("background.png")
|
20 |
+
font = ImageFont.truetype("Lato-Regular.ttf", 40)
|
21 |
+
text_color = (255, 200, 200)
|
22 |
+
highlight_color = (255, 255, 255)
|
23 |
+
|
24 |
+
# checkpoint = "openai/whisper-tiny"
|
25 |
+
# checkpoint = "openai/whisper-base"
|
26 |
+
checkpoint = "openai/whisper-small"
|
27 |
+
pipe = pipeline(model=checkpoint)
|
28 |
+
|
29 |
+
# TODO: no longer need to set these manually once the models have been updated on the Hub
|
30 |
+
# whisper-base
|
31 |
+
# pipe.model.config.alignment_heads = [[3, 1], [4, 2], [4, 3], [4, 7], [5, 1], [5, 2], [5, 4], [5, 6]]
|
32 |
+
# whisper-small
|
33 |
+
pipe.model.config.alignment_heads = [[5, 3], [5, 9], [8, 0], [8, 4], [8, 7], [8, 8], [9, 0], [9, 7], [9, 9], [10, 5]]
|
34 |
+
|
35 |
+
chunks = []
|
36 |
+
|
37 |
+
|
38 |
+
def make_frame(t):
|
39 |
+
global chunks
|
40 |
+
|
41 |
+
# TODO speed optimization: could cache the last image returned and if the
|
42 |
+
# active chunk and active word didn't change, use that last image instead
|
43 |
+
# of drawing the exact same thing again
|
44 |
+
|
45 |
+
# TODO in the Henry V example, the word "desires" has an ending timestamp
|
46 |
+
# that's too far into the future, and so the word stays highlighted.
|
47 |
+
# Could fix this by finding the latest word that is active in the chunk
|
48 |
+
# and only highlight that one.
|
49 |
+
|
50 |
+
image = background_image.copy()
|
51 |
+
draw = ImageDraw.Draw(image)
|
52 |
+
|
53 |
+
# for debugging: draw frame time
|
54 |
+
#draw.text((20, 20), str(t), fill=text_color, font=font)
|
55 |
+
|
56 |
+
space_length = draw.textlength(" ", font)
|
57 |
+
x = margin_left
|
58 |
+
y = margin_top
|
59 |
+
|
60 |
+
for chunk in chunks:
|
61 |
+
chunk_start = chunk["timestamp"][0]
|
62 |
+
chunk_end = chunk["timestamp"][1]
|
63 |
+
if chunk_end is None: chunk_end = max_duration
|
64 |
+
|
65 |
+
if chunk_start <= t <= chunk_end:
|
66 |
+
words = [x["text"] for x in chunk["words"]]
|
67 |
+
word_times = [x["timestamp"] for x in chunk["words"]]
|
68 |
+
|
69 |
+
for (word, times) in zip(words, word_times):
|
70 |
+
word_length = draw.textlength(word + " ", font) - space_length
|
71 |
+
if x + word_length >= video_width - margin_right:
|
72 |
+
x = margin_left
|
73 |
+
y += line_height
|
74 |
+
|
75 |
+
if times[0] <= t <= times[1]:
|
76 |
+
color = highlight_color
|
77 |
+
draw.rectangle([x, y + line_height, x + word_length, y + line_height + 4], fill=color)
|
78 |
+
else:
|
79 |
+
color = text_color
|
80 |
+
|
81 |
+
draw.text((x, y), word, fill=color, font=font)
|
82 |
+
x += word_length + space_length
|
83 |
+
|
84 |
+
break
|
85 |
+
|
86 |
+
return np.array(image)
|
87 |
+
|
88 |
+
|
89 |
+
def predict(audio_path):
|
90 |
+
global chunks
|
91 |
+
|
92 |
+
audio_data, sr = librosa.load(audio_path, mono=True)
|
93 |
+
duration = librosa.get_duration(y=audio_data, sr=sr)
|
94 |
+
duration = min(max_duration, duration)
|
95 |
+
audio_data = audio_data[:int(duration * sr)]
|
96 |
+
|
97 |
+
# Run Whisper to get word-level timestamps.
|
98 |
+
audio_inputs = librosa.resample(audio_data, orig_sr=sr, target_sr=pipe.feature_extractor.sampling_rate)
|
99 |
+
output = pipe(audio_inputs, chunk_length_s=30, stride_length_s=[4, 2], return_timestamps="word")
|
100 |
+
chunks = output["chunks"]
|
101 |
+
print(chunks)
|
102 |
+
|
103 |
+
# Create the video.
|
104 |
+
clip = mpy.VideoClip(make_frame, duration=duration)
|
105 |
+
audio_clip = mpy.AudioFileClip(audio_path).set_duration(duration)
|
106 |
+
clip = clip.set_audio(audio_clip)
|
107 |
+
clip.write_videofile("my_video.mp4", fps=fps, codec="libx264", audio_codec="aac")
|
108 |
+
return "my_video.mp4"
|
109 |
+
|
110 |
+
|
111 |
+
title = "Word-level timestamps with Whisper"
|
112 |
+
|
113 |
+
description = """
|
114 |
+
This demo shows Whisper <b>word-level timestamps</b> in action using Hugging Face Transformers. It creates a video showing subtitled audio with the current word highlighted.
|
115 |
+
|
116 |
+
This demo uses the <b>openai/whisper-small</b> checkpoint. Since it's only a demo, the output is limited to the first 60 seconds of audio.
|
117 |
+
"""
|
118 |
+
|
119 |
+
article = """
|
120 |
+
<div style='margin:20px auto;'>
|
121 |
+
|
122 |
+
<p>Credits:<p>
|
123 |
+
|
124 |
+
<ul>
|
125 |
+
<li>Shakespeare's "Henry V" speech from <a href="https://freesound.org/people/acclivity/sounds/24096/">acclivity</a> (CC BY-NC 4.0 license)
|
126 |
+
<li>Lato font by Łukasz Dziedzic (licensed under Open Font License)</li>
|
127 |
+
<li>Whisper model by OpenAI</li>
|
128 |
+
</ul>
|
129 |
+
|
130 |
+
</div>
|
131 |
+
"""
|
132 |
+
|
133 |
+
examples = [
|
134 |
+
"examples/henry5.wav",
|
135 |
+
]
|
136 |
+
|
137 |
+
gr.Interface(
|
138 |
+
fn=predict,
|
139 |
+
inputs=[
|
140 |
+
gr.Audio(label="Upload Audio", source="upload", type="filepath"),
|
141 |
+
],
|
142 |
+
outputs=[
|
143 |
+
gr.Video(label="Output Video"),
|
144 |
+
],
|
145 |
+
title=title,
|
146 |
+
description=description,
|
147 |
+
article=article,
|
148 |
+
examples=examples,
|
149 |
+
).launch()
|
background.png
ADDED
requirements.txt
ADDED
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
git+https://github.com/hollance/transformers.git@whisper_word_timestamps
|
2 |
+
torch
|
3 |
+
torchaudio
|
4 |
+
soundfile
|
5 |
+
librosa
|
6 |
+
moviepy
|
7 |
+
matplotlib
|
8 |
+
pillow
|