Upload app.py
Browse files
app.py
ADDED
@@ -0,0 +1,122 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
|
2 |
+
import gradio as gr
|
3 |
+
import random
|
4 |
+
import json
|
5 |
+
import os
|
6 |
+
from difflib import SequenceMatcher
|
7 |
+
from jiwer import wer
|
8 |
+
import torchaudio
|
9 |
+
import torch
|
10 |
+
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor, HubertForCTC, HubertProcessor
|
11 |
+
import whisper
|
12 |
+
|
13 |
+
# Load metadata
|
14 |
+
with open("common_voice_en_validated_249_hf_ready.json") as f:
|
15 |
+
data = json.load(f)
|
16 |
+
|
17 |
+
# Available filter values
|
18 |
+
ages = sorted(set(entry["age"] for entry in data))
|
19 |
+
genders = sorted(set(entry["gender"] for entry in data))
|
20 |
+
accents = sorted(set(entry["accent"] for entry in data))
|
21 |
+
|
22 |
+
# Load models
|
23 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
24 |
+
|
25 |
+
# Whisper
|
26 |
+
whisper_model = whisper.load_model("medium").to(device)
|
27 |
+
|
28 |
+
# Wav2Vec2
|
29 |
+
wav2vec_processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-large-960h-lv60-self")
|
30 |
+
wav2vec_model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-large-960h-lv60-self").to(device)
|
31 |
+
|
32 |
+
# HuBERT
|
33 |
+
hubert_processor = HubertProcessor.from_pretrained("facebook/hubert-large-ls960-ft")
|
34 |
+
hubert_model = HubertForCTC.from_pretrained("facebook/hubert-large-ls960-ft").to(device)
|
35 |
+
|
36 |
+
def load_audio(file_path):
|
37 |
+
waveform, sr = torchaudio.load(file_path)
|
38 |
+
return torchaudio.functional.resample(waveform, orig_freq=sr, new_freq=16000)[0].numpy()
|
39 |
+
|
40 |
+
def transcribe_whisper(file_path):
|
41 |
+
result = whisper_model.transcribe(file_path)
|
42 |
+
return result["text"].strip().lower()
|
43 |
+
|
44 |
+
def transcribe_wav2vec(file_path):
|
45 |
+
audio = load_audio(file_path)
|
46 |
+
inputs = wav2vec_processor(audio, sampling_rate=16000, return_tensors="pt", padding=True)
|
47 |
+
with torch.no_grad():
|
48 |
+
logits = wav2vec_model(**inputs.to(device)).logits
|
49 |
+
predicted_ids = torch.argmax(logits, dim=-1)
|
50 |
+
return wav2vec_processor.batch_decode(predicted_ids)[0].strip().lower()
|
51 |
+
|
52 |
+
def transcribe_hubert(file_path):
|
53 |
+
audio = load_audio(file_path)
|
54 |
+
inputs = hubert_processor(audio, sampling_rate=16000, return_tensors="pt", padding=True)
|
55 |
+
with torch.no_grad():
|
56 |
+
logits = hubert_model(**inputs.to(device)).logits
|
57 |
+
predicted_ids = torch.argmax(logits, dim=-1)
|
58 |
+
return hubert_processor.batch_decode(predicted_ids)[0].strip().lower()
|
59 |
+
|
60 |
+
def highlight_differences(ref, hyp):
|
61 |
+
sm = SequenceMatcher(None, ref.split(), hyp.split())
|
62 |
+
result = []
|
63 |
+
for opcode, i1, i2, j1, j2 in sm.get_opcodes():
|
64 |
+
if opcode == 'equal':
|
65 |
+
result.extend(hyp.split()[j1:j2])
|
66 |
+
elif opcode in ('replace', 'insert', 'delete'):
|
67 |
+
wrong = hyp.split()[j1:j2]
|
68 |
+
result.extend([f"<span style='color:red'>{w}</span>" for w in wrong])
|
69 |
+
return " ".join(result)
|
70 |
+
|
71 |
+
def run_demo(age, gender, accent):
|
72 |
+
filtered = [
|
73 |
+
entry for entry in data
|
74 |
+
if entry["age"] == age and entry["gender"] == gender and entry["accent"] == accent
|
75 |
+
]
|
76 |
+
if not filtered:
|
77 |
+
return "No matching sample.", None, "", "", "", "", "", ""
|
78 |
+
|
79 |
+
sample = random.choice(filtered)
|
80 |
+
file_path = os.path.join("common_voice_en_validated_249", sample["path"])
|
81 |
+
gold = sample["sentence"].strip().lower()
|
82 |
+
|
83 |
+
whisper_text = transcribe_whisper(file_path)
|
84 |
+
wav2vec_text = transcribe_wav2vec(file_path)
|
85 |
+
hubert_text = transcribe_hubert(file_path)
|
86 |
+
|
87 |
+
table = f"""
|
88 |
+
<table border="1" style="width:100%">
|
89 |
+
<tr><th>Model</th><th>Transcription</th><th>WER</th></tr>
|
90 |
+
<tr><td><b>Gold</b></td><td>{gold}</td><td>0.00</td></tr>
|
91 |
+
<tr><td>Whisper</td><td>{highlight_differences(gold, whisper_text)}</td><td>{wer(gold, whisper_text):.2f}</td></tr>
|
92 |
+
<tr><td>Wav2Vec2</td><td>{highlight_differences(gold, wav2vec_text)}</td><td>{wer(gold, wav2vec_text):.2f}</td></tr>
|
93 |
+
<tr><td>HuBERT</td><td>{highlight_differences(gold, hubert_text)}</td><td>{wer(gold, hubert_text):.2f}</td></tr>
|
94 |
+
</table>
|
95 |
+
"""
|
96 |
+
|
97 |
+
return sample["sentence"], file_path, gold, whisper_text, wav2vec_text, hubert_text, table, f"Audio path: {file_path}"
|
98 |
+
|
99 |
+
with gr.Blocks() as demo:
|
100 |
+
gr.Markdown("# ASR Model Comparison on ESL Audio")
|
101 |
+
gr.Markdown("Filter by age, gender, and accent. Then generate a random ESL learner's audio to compare how Whisper, Wav2Vec2, and HuBERT transcribe it.")
|
102 |
+
|
103 |
+
with gr.Row():
|
104 |
+
age = gr.Dropdown(choices=ages, label="Age")
|
105 |
+
gender = gr.Dropdown(choices=genders, label="Gender")
|
106 |
+
accent = gr.Dropdown(choices=accents, label="Accent")
|
107 |
+
|
108 |
+
btn = gr.Button("Generate and Transcribe")
|
109 |
+
audio = gr.Audio(label="Audio", type="filepath")
|
110 |
+
wer_output = gr.HTML()
|
111 |
+
|
112 |
+
btn.click(fn=run_demo, inputs=[age, gender, accent], outputs=[
|
113 |
+
gr.Textbox(label="Gold (Correct)"),
|
114 |
+
audio,
|
115 |
+
gr.Textbox(label="Whisper Output"),
|
116 |
+
gr.Textbox(label="Wav2Vec2 Output"),
|
117 |
+
gr.Textbox(label="HuBERT Output"),
|
118 |
+
wer_output,
|
119 |
+
gr.Textbox(label="Path")
|
120 |
+
])
|
121 |
+
|
122 |
+
demo.launch()
|