Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -1,8 +1,6 @@
|
|
1 |
import gradio as gr
|
2 |
-
import nemo.collections.asr as nemo_asr
|
3 |
from nemo.core import ModelPT
|
4 |
import torch
|
5 |
-
from evaluate import load as hf_load
|
6 |
import os
|
7 |
import spaces
|
8 |
|
@@ -19,9 +17,6 @@ MODEL_NAMES = [
|
|
19 |
# Cache loaded models
|
20 |
LOADED_MODELS = {}
|
21 |
|
22 |
-
# Load WER and CER metrics from HuggingFace evaluate
|
23 |
-
hf_wer = hf_load("wer")
|
24 |
-
hf_cer = hf_load("cer")
|
25 |
|
26 |
def get_model(model_name):
|
27 |
if model_name not in LOADED_MODELS:
|
@@ -38,127 +33,56 @@ def get_model(model_name):
|
|
38 |
return LOADED_MODELS[model_name]
|
39 |
|
40 |
@spaces.GPU(duration=120)
|
41 |
-
def transcribe_and_score(audio
|
42 |
-
if audio is None
|
43 |
-
return ""
|
44 |
-
model = get_model(
|
45 |
|
46 |
# Use the correct transcribe API
|
47 |
predictions = model.transcribe([audio])
|
48 |
pred = predictions[0] if isinstance(predictions, list) else predictions
|
49 |
|
50 |
-
# Ensure both are strings and not empty
|
51 |
-
if not isinstance(ground_truth, str):
|
52 |
-
ground_truth = str(ground_truth)
|
53 |
if not isinstance(pred, str):
|
54 |
pred = str(pred)
|
55 |
|
56 |
-
|
57 |
-
ground_truth = ground_truth.strip()
|
58 |
-
pred = pred.strip()
|
59 |
-
|
60 |
-
# Debug output
|
61 |
-
print(f"[DEBUG] Model: {model_name}")
|
62 |
-
print(f"[DEBUG] Ground truth: '{ground_truth}' (length: {len(ground_truth)})")
|
63 |
-
print(f"[DEBUG] Prediction: '{pred}' (length: {len(pred)})")
|
64 |
-
print(f"[DEBUG] Are they equal? {ground_truth == pred}")
|
65 |
-
print(f"[DEBUG] Ground truth bytes: {repr(ground_truth)}")
|
66 |
-
print(f"[DEBUG] Prediction bytes: {repr(pred)}")
|
67 |
-
|
68 |
-
if not ground_truth or not pred:
|
69 |
-
print("[DEBUG] Empty ground truth or prediction, returning 1.0")
|
70 |
-
return pred, 1.0, 1.0
|
71 |
-
|
72 |
-
# Calculate WER and CER
|
73 |
-
wer_score = hf_wer.compute(predictions=[pred], references=[ground_truth])
|
74 |
-
cer_score = hf_cer.compute(predictions=[pred], references=[ground_truth])
|
75 |
-
|
76 |
-
print(f"[DEBUG] WER: {wer_score}, CER: {cer_score}")
|
77 |
-
return pred, wer_score, cer_score
|
78 |
|
79 |
@spaces.GPU(duration=120)
|
80 |
-
def
|
81 |
-
if not audio_files
|
82 |
-
return []
|
83 |
-
|
84 |
-
|
85 |
-
|
86 |
-
model = get_model(model_name)
|
87 |
-
|
88 |
# Use the correct transcribe API for batch
|
89 |
predictions = model.transcribe(audio_files)
|
90 |
-
|
91 |
-
|
92 |
-
|
93 |
-
|
94 |
-
for i, (audio_file, gt) in enumerate(zip(audio_files, ground_truths)):
|
95 |
-
pred = predictions[i] if isinstance(predictions, list) else predictions
|
96 |
-
|
97 |
-
if not isinstance(gt, str):
|
98 |
-
gt = str(gt)
|
99 |
-
if not isinstance(pred, str):
|
100 |
-
pred = str(pred)
|
101 |
-
|
102 |
-
gt = gt.strip()
|
103 |
-
pred = pred.strip()
|
104 |
-
|
105 |
-
if not gt or not pred:
|
106 |
-
wer_score = 1.0
|
107 |
-
cer_score = 1.0
|
108 |
-
else:
|
109 |
-
wer_score = hf_wer.compute(predictions=[pred], references=[gt])
|
110 |
-
cer_score = hf_cer.compute(predictions=[pred], references=[gt])
|
111 |
-
|
112 |
-
results.append([pred, wer_score, cer_score])
|
113 |
-
pred_texts.append(pred)
|
114 |
-
|
115 |
-
# Calculate average WER and CER
|
116 |
-
if pred_texts and ground_truths:
|
117 |
-
avg_wer = hf_wer.compute(predictions=pred_texts, references=ground_truths)
|
118 |
-
avg_cer = hf_cer.compute(predictions=pred_texts, references=ground_truths)
|
119 |
else:
|
120 |
-
|
121 |
-
|
122 |
-
|
123 |
-
return
|
124 |
|
125 |
with gr.Blocks(title="EgypTalk-ASR-v2") as demo:
|
126 |
gr.Markdown("""
|
127 |
# EgypTalk-ASR-v2
|
128 |
-
Upload an audio file
|
129 |
""")
|
130 |
with gr.Tab("Single Test"):
|
131 |
with gr.Row():
|
132 |
audio_input = gr.Audio(type="filepath", label="Audio File")
|
133 |
-
|
134 |
-
|
135 |
-
transcribe_btn
|
136 |
-
with gr.Row():
|
137 |
-
pred_output = gr.Textbox(label="Transcription")
|
138 |
-
wer_output = gr.Number(label="WER")
|
139 |
-
cer_output = gr.Number(label="CER")
|
140 |
-
transcribe_btn.click(transcribe_and_score, inputs=[audio_input, gt_input, model_choice], outputs=[pred_output, wer_output, cer_output])
|
141 |
|
142 |
with gr.Tab("Batch Test"):
|
143 |
-
gr.Markdown("Upload multiple audio files
|
144 |
audio_files = gr.Files(label="Audio Files (wav)")
|
145 |
-
|
146 |
-
|
147 |
-
batch_btn = gr.Button("Batch Transcribe & Evaluate")
|
148 |
-
preds_output = gr.Dataframe(headers=["Prediction", "WER", "CER"], label="Results")
|
149 |
-
avg_wer_output = gr.Number(label="Average WER")
|
150 |
-
avg_cer_output = gr.Number(label="Average CER")
|
151 |
-
|
152 |
-
def batch_wrapper(audio_files, gt_file, model_name):
|
153 |
-
if not audio_files or not gt_file:
|
154 |
-
return [], None, None
|
155 |
-
with open(gt_file, 'r', encoding='utf-8') as f:
|
156 |
-
gts = [line.strip() for line in f if line.strip()]
|
157 |
-
audio_files_sorted = sorted(audio_files, key=lambda x: os.path.basename(x))
|
158 |
-
results, avg_wer, avg_cer = batch_transcribe_and_score(audio_files_sorted, gts, model_name)
|
159 |
-
return results, avg_wer, avg_cer
|
160 |
|
161 |
-
batch_btn.click(
|
162 |
|
163 |
|
164 |
demo.launch(share=True)
|
|
|
1 |
import gradio as gr
|
|
|
2 |
from nemo.core import ModelPT
|
3 |
import torch
|
|
|
4 |
import os
|
5 |
import spaces
|
6 |
|
|
|
17 |
# Cache loaded models
|
18 |
LOADED_MODELS = {}
|
19 |
|
|
|
|
|
|
|
20 |
|
21 |
def get_model(model_name):
|
22 |
if model_name not in LOADED_MODELS:
|
|
|
33 |
return LOADED_MODELS[model_name]
|
34 |
|
35 |
@spaces.GPU(duration=120)
|
36 |
+
def transcribe_and_score(audio):
|
37 |
+
if audio is None:
|
38 |
+
return ""
|
39 |
+
model = get_model(MODEL_NAMES[0])
|
40 |
|
41 |
# Use the correct transcribe API
|
42 |
predictions = model.transcribe([audio])
|
43 |
pred = predictions[0] if isinstance(predictions, list) else predictions
|
44 |
|
|
|
|
|
|
|
45 |
if not isinstance(pred, str):
|
46 |
pred = str(pred)
|
47 |
|
48 |
+
return pred.strip()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
49 |
|
50 |
@spaces.GPU(duration=120)
|
51 |
+
def batch_transcribe(audio_files):
|
52 |
+
if not audio_files:
|
53 |
+
return []
|
54 |
+
model = get_model(MODEL_NAMES[0])
|
55 |
+
|
|
|
|
|
|
|
56 |
# Use the correct transcribe API for batch
|
57 |
predictions = model.transcribe(audio_files)
|
58 |
+
|
59 |
+
if isinstance(predictions, list):
|
60 |
+
texts = [p if isinstance(p, str) else str(p) for p in predictions]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
61 |
else:
|
62 |
+
texts = [str(predictions)]
|
63 |
+
|
64 |
+
# Return as rows for a single-column dataframe
|
65 |
+
return [[t.strip()] for t in texts]
|
66 |
|
67 |
with gr.Blocks(title="EgypTalk-ASR-v2") as demo:
|
68 |
gr.Markdown("""
|
69 |
# EgypTalk-ASR-v2
|
70 |
+
Upload an audio file. This app transcribes audio using EgypTalk-ASR-v2.
|
71 |
""")
|
72 |
with gr.Tab("Single Test"):
|
73 |
with gr.Row():
|
74 |
audio_input = gr.Audio(type="filepath", label="Audio File")
|
75 |
+
transcribe_btn = gr.Button("Transcribe")
|
76 |
+
pred_output = gr.Textbox(label="Transcription")
|
77 |
+
transcribe_btn.click(transcribe_and_score, inputs=[audio_input], outputs=[pred_output])
|
|
|
|
|
|
|
|
|
|
|
78 |
|
79 |
with gr.Tab("Batch Test"):
|
80 |
+
gr.Markdown("Upload multiple audio files. Batch size is limited by GPU/CPU memory.")
|
81 |
audio_files = gr.Files(label="Audio Files (wav)")
|
82 |
+
batch_btn = gr.Button("Batch Transcribe")
|
83 |
+
preds_output = gr.Dataframe(headers=["Transcription"], label="Results")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
84 |
|
85 |
+
batch_btn.click(batch_transcribe, inputs=[audio_files], outputs=[preds_output])
|
86 |
|
87 |
|
88 |
demo.launch(share=True)
|