Spaces:
Sleeping
Sleeping
update model
Browse files
app.py
CHANGED
@@ -5,7 +5,6 @@ import tempfile
|
|
5 |
import requests
|
6 |
from moviepy.editor import VideoFileClip
|
7 |
|
8 |
-
# Ensure the official OpenAI Whisper package is installed (supports load_model)
|
9 |
try:
|
10 |
import whisper
|
11 |
if not hasattr(whisper, 'load_model'):
|
@@ -21,29 +20,26 @@ from transformers import Wav2Vec2Processor, Wav2Vec2ForSequenceClassification
|
|
21 |
from huggingface_hub import login
|
22 |
import gradio as gr
|
23 |
|
24 |
-
# Authenticate with Hugging Face (token via HF_TOKEN env var)
|
25 |
|
26 |
-
|
27 |
-
|
28 |
-
# Device setup (GPU if available)
|
29 |
-
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
30 |
|
31 |
def load_models():
|
32 |
-
|
33 |
-
whisper_model = whisper.load_model('base', device=device)
|
34 |
processor = Wav2Vec2Processor.from_pretrained(
|
35 |
-
'jonatasgrosman/wav2vec2-large-
|
36 |
)
|
37 |
accent_model = Wav2Vec2ForSequenceClassification.from_pretrained(
|
38 |
-
'jonatasgrosman/wav2vec2-large-
|
39 |
).to(device)
|
|
|
|
|
|
|
40 |
return whisper_model, processor, accent_model
|
41 |
|
42 |
whisper_model, processor, accent_model = load_models()
|
43 |
|
44 |
-
# Main analysis function
|
45 |
def analyze(video_url: str):
|
46 |
-
# Download video to temp file
|
47 |
with tempfile.NamedTemporaryFile(delete=False, suffix='.mp4') as tmp_vid:
|
48 |
response = requests.get(video_url, stream=True)
|
49 |
response.raise_for_status()
|
@@ -52,23 +48,19 @@ def analyze(video_url: str):
|
|
52 |
tmp_vid.write(chunk)
|
53 |
video_path = tmp_vid.name
|
54 |
|
55 |
-
# Extract audio
|
56 |
audio_path = video_path.replace('.mp4', '.wav')
|
57 |
clip = VideoFileClip(video_path)
|
58 |
clip.audio.write_audiofile(audio_path, verbose=False, logger=None)
|
59 |
clip.close()
|
60 |
|
61 |
-
# Load audio waveform
|
62 |
speech, sr = librosa.load(audio_path, sr=16000)
|
63 |
|
64 |
-
# Transcribe with Whisper (model on correct device)
|
65 |
result = whisper_model.transcribe(speech)
|
66 |
transcript = result.get('text', '')
|
67 |
lang = result.get('language', 'unknown')
|
68 |
if lang != 'en':
|
69 |
transcript = f"[Non-English detected: {lang}]\n" + transcript
|
70 |
|
71 |
-
# Accent classification
|
72 |
inputs = processor(speech, sampling_rate=sr, return_tensors='pt', padding=True)
|
73 |
input_values = inputs.input_values.to(device)
|
74 |
attention_mask = inputs.attention_mask.to(device)
|
@@ -76,20 +68,17 @@ def analyze(video_url: str):
|
|
76 |
logits = accent_model(input_values=input_values, attention_mask=attention_mask).logits
|
77 |
probs = torch.softmax(logits, dim=-1).squeeze().cpu().tolist()
|
78 |
|
79 |
-
# Map default LABEL_x to human-readable accents
|
80 |
accent_labels = [
|
81 |
'American', 'Australian', 'British', 'Canadian', 'Indian',
|
82 |
'Irish', 'New Zealander', 'South African', 'Welsh'
|
83 |
-
]
|
84 |
accent_probs = [(accent_labels[i], probs[i] * 100) for i in range(len(probs))]
|
85 |
accent_probs.sort(key=lambda x: x[1], reverse=True)
|
86 |
top_accent, top_conf = accent_probs[0]
|
87 |
|
88 |
-
# Prepare DataFrame
|
89 |
df = pd.DataFrame(accent_probs, columns=['Accent', 'Confidence (%)'])
|
90 |
df = pd.DataFrame(accent_probs, columns=['Accent', 'Confidence (%)'])
|
91 |
|
92 |
-
# Cleanup temp files
|
93 |
try:
|
94 |
os.remove(video_path)
|
95 |
os.remove(audio_path)
|
@@ -98,7 +87,6 @@ def analyze(video_url: str):
|
|
98 |
|
99 |
return top_accent, f"{top_conf:.2f}%", df
|
100 |
|
101 |
-
# Gradio interface
|
102 |
interface = gr.Interface(
|
103 |
fn=analyze,
|
104 |
inputs=gr.Textbox(label='Video URL', placeholder='Enter public MP4 URL'),
|
@@ -109,7 +97,7 @@ interface = gr.Interface(
|
|
109 |
gr.Dataframe(label='All Accent Probabilities')
|
110 |
],
|
111 |
title='English Accent Detector',
|
112 |
-
description='Paste a
|
113 |
allow_flagging='never'
|
114 |
)
|
115 |
|
|
|
5 |
import requests
|
6 |
from moviepy.editor import VideoFileClip
|
7 |
|
|
|
8 |
try:
|
9 |
import whisper
|
10 |
if not hasattr(whisper, 'load_model'):
|
|
|
20 |
from huggingface_hub import login
|
21 |
import gradio as gr
|
22 |
|
|
|
23 |
|
24 |
+
# device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
25 |
+
device = 'cpu'
|
|
|
|
|
26 |
|
27 |
def load_models():
|
28 |
+
whisper_model = whisper.load_model('tiny', device=device)
|
|
|
29 |
processor = Wav2Vec2Processor.from_pretrained(
|
30 |
+
'jonatasgrosman/wav2vec2-large-english'
|
31 |
)
|
32 |
accent_model = Wav2Vec2ForSequenceClassification.from_pretrained(
|
33 |
+
'jonatasgrosman/wav2vec2-large-english'
|
34 |
).to(device)
|
35 |
+
accent_model = torch.quantization.quantize_dynamic(
|
36 |
+
accent_model, {torch.nn.Linear}, dtype=torch.qint8
|
37 |
+
)
|
38 |
return whisper_model, processor, accent_model
|
39 |
|
40 |
whisper_model, processor, accent_model = load_models()
|
41 |
|
|
|
42 |
def analyze(video_url: str):
|
|
|
43 |
with tempfile.NamedTemporaryFile(delete=False, suffix='.mp4') as tmp_vid:
|
44 |
response = requests.get(video_url, stream=True)
|
45 |
response.raise_for_status()
|
|
|
48 |
tmp_vid.write(chunk)
|
49 |
video_path = tmp_vid.name
|
50 |
|
|
|
51 |
audio_path = video_path.replace('.mp4', '.wav')
|
52 |
clip = VideoFileClip(video_path)
|
53 |
clip.audio.write_audiofile(audio_path, verbose=False, logger=None)
|
54 |
clip.close()
|
55 |
|
|
|
56 |
speech, sr = librosa.load(audio_path, sr=16000)
|
57 |
|
|
|
58 |
result = whisper_model.transcribe(speech)
|
59 |
transcript = result.get('text', '')
|
60 |
lang = result.get('language', 'unknown')
|
61 |
if lang != 'en':
|
62 |
transcript = f"[Non-English detected: {lang}]\n" + transcript
|
63 |
|
|
|
64 |
inputs = processor(speech, sampling_rate=sr, return_tensors='pt', padding=True)
|
65 |
input_values = inputs.input_values.to(device)
|
66 |
attention_mask = inputs.attention_mask.to(device)
|
|
|
68 |
logits = accent_model(input_values=input_values, attention_mask=attention_mask).logits
|
69 |
probs = torch.softmax(logits, dim=-1).squeeze().cpu().tolist()
|
70 |
|
|
|
71 |
accent_labels = [
|
72 |
'American', 'Australian', 'British', 'Canadian', 'Indian',
|
73 |
'Irish', 'New Zealander', 'South African', 'Welsh'
|
74 |
+
]
|
75 |
accent_probs = [(accent_labels[i], probs[i] * 100) for i in range(len(probs))]
|
76 |
accent_probs.sort(key=lambda x: x[1], reverse=True)
|
77 |
top_accent, top_conf = accent_probs[0]
|
78 |
|
|
|
79 |
df = pd.DataFrame(accent_probs, columns=['Accent', 'Confidence (%)'])
|
80 |
df = pd.DataFrame(accent_probs, columns=['Accent', 'Confidence (%)'])
|
81 |
|
|
|
82 |
try:
|
83 |
os.remove(video_path)
|
84 |
os.remove(audio_path)
|
|
|
87 |
|
88 |
return top_accent, f"{top_conf:.2f}%", df
|
89 |
|
|
|
90 |
interface = gr.Interface(
|
91 |
fn=analyze,
|
92 |
inputs=gr.Textbox(label='Video URL', placeholder='Enter public MP4 URL'),
|
|
|
97 |
gr.Dataframe(label='All Accent Probabilities')
|
98 |
],
|
99 |
title='English Accent Detector',
|
100 |
+
description='Paste a direct MP4 URL to extract, transcribe, and classify English accents. It is a bit slow since we run Whisper and Wav2Vec2 models on CPU. Please test with short videos.',
|
101 |
allow_flagging='never'
|
102 |
)
|
103 |
|