Owen
commited on
Commit
·
c6818dd
1
Parent(s):
e909f31
add conformer
Browse files- .gitattributes +4 -2
- app.py +114 -16
- jawa.wav +3 -0
- requirements.txt +4 -1
- sunda.wav +3 -0
- test.py +5 -0
.gitattributes
CHANGED
@@ -33,5 +33,7 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
36 |
-
conformer.png filter=lfs diff=lfs merge=lfs -text
|
37 |
-
whisper.png filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
36 |
+
conformer.png filter=lfs diff=lfs merge=lfs -text
|
37 |
+
whisper.png filter=lfs diff=lfs merge=lfs -text
|
38 |
+
*.wav filter=lfs diff=lfs merge=lfs -text
|
39 |
+
*.png filter=lfs diff=lfs merge=lfs -text
|
app.py
CHANGED
@@ -1,12 +1,76 @@
|
|
|
|
|
|
|
|
|
|
1 |
import numpy as np # type: ignore
|
2 |
import gradio as gr # type: ignore
|
3 |
from transformers import pipeline
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
4 |
|
5 |
# Load fine-tuned Whisper model
|
6 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
7 |
|
8 |
-
def transcribe(audio):
|
9 |
sr, waveform = audio
|
|
|
10 |
# Change into Mono Audio
|
11 |
if waveform.ndim > 1:
|
12 |
waveform = waveform.mean(axis=1)
|
@@ -15,7 +79,33 @@ def transcribe(audio):
|
|
15 |
waveform = waveform.astype(np.float32)
|
16 |
waveform /= np.max(np.abs(waveform))
|
17 |
|
18 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
19 |
"sampling_rate" : sr,
|
20 |
"raw" : waveform
|
21 |
})["text"]
|
@@ -25,6 +115,7 @@ def clear():
|
|
25 |
|
26 |
# --- Tab 1: Transcribe ---
|
27 |
with gr.Blocks() as tab_transcribe:
|
|
|
28 |
with gr.Row():
|
29 |
with gr.Column(scale=1):
|
30 |
audio_input = gr.Audio(sources="microphone", label="Record Your Voice")
|
@@ -35,7 +126,13 @@ with gr.Blocks() as tab_transcribe:
|
|
35 |
with gr.Column(scale=1):
|
36 |
output_text = gr.Textbox(label="Transcription", placeholder="Waiting for Input", lines=3)
|
37 |
|
38 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
39 |
clrBtn.click(fn=clear, outputs=[audio_input, output_text])
|
40 |
|
41 |
# --- Tab 2: Penjelasan Model Fine-Tuned ---
|
@@ -52,19 +149,20 @@ with gr.Blocks() as tab_background:
|
|
52 |
Model yang telah kami fine tune merupakan hasil <b>fine-tuning dari Whisper dan Conformer</b> untuk mendukung bahasa lokal di Indonesia, khususnya bahasa Jawa dan Sunda.
|
53 |
Model dilatih menggunakan kombinasi dataset <b>OpenSLR</b> berikut:
|
54 |
<br>
|
55 |
-
|
56 |
-
|
57 |
-
|
58 |
-
|
59 |
<br>
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
|
|
|
|
|
|
64 |
<br>
|
65 |
-
<a href="https://openslr.org/
|
66 |
-
<b>
|
67 |
-
<b>SLR44</b> - Bilingual speech datasets
|
68 |
</a>
|
69 |
|
70 |
<br>
|
@@ -172,7 +270,7 @@ demo = gr.TabbedInterface(
|
|
172 |
[tab_transcribe, tab_background, tab_architecture, tab_results, tab_authors],
|
173 |
["Transcribe", "Latar Belakang", "Arsitektur", "Evaluasi", "Fine-Tuned By"],
|
174 |
theme=gr.themes.Soft(),
|
175 |
-
title="
|
176 |
)
|
177 |
|
178 |
if __name__ == "__main__":
|
|
|
1 |
+
import os
|
2 |
+
import torch
|
3 |
+
import torch.nn as nn
|
4 |
+
import torchaudio
|
5 |
import numpy as np # type: ignore
|
6 |
import gradio as gr # type: ignore
|
7 |
from transformers import pipeline
|
8 |
+
from huggingface_hub import hf_hub_download
|
9 |
+
from torchaudio.models import Conformer
|
10 |
+
|
11 |
+
class ASRConformerModel(nn.Module):
|
12 |
+
def __init__(self, input_dim, vocab_size):
|
13 |
+
super().__init__()
|
14 |
+
self.encoder = Conformer(
|
15 |
+
input_dim=input_dim,
|
16 |
+
num_heads=4,
|
17 |
+
ffn_dim=512,
|
18 |
+
num_layers=4,
|
19 |
+
depthwise_conv_kernel_size=31,
|
20 |
+
dropout=0.1
|
21 |
+
)
|
22 |
+
self.classifier = nn.Linear(input_dim, vocab_size)
|
23 |
+
|
24 |
+
def forward(self, x, lengths):
|
25 |
+
x, lengths = self.encoder(x, lengths=lengths)
|
26 |
+
x = self.classifier(x)
|
27 |
+
return x, lengths
|
28 |
+
|
29 |
+
|
30 |
+
VOCAB = set("abcdefghijklmnopqrstuvwxyz '")
|
31 |
+
char_to_idx = {ch: i + 1 for i, ch in enumerate(sorted(VOCAB))} # 0 for CTC blank
|
32 |
+
|
33 |
+
def greedy_decode(log_probs, blank=0):
|
34 |
+
pred_ids = log_probs.argmax(dim=-1) # [T, B]
|
35 |
+
pred_ids = pred_ids.transpose(0, 1) # [B, T]
|
36 |
+
predictions = []
|
37 |
+
for seq in pred_ids:
|
38 |
+
prev = blank
|
39 |
+
pred = []
|
40 |
+
for i in seq:
|
41 |
+
if i != prev and i != blank:
|
42 |
+
pred.append(i.item())
|
43 |
+
prev = i
|
44 |
+
predictions.append(pred)
|
45 |
+
return predictions
|
46 |
+
|
47 |
+
def encode(text):
|
48 |
+
return torch.tensor([char_to_idx[c] for c in text.lower() if c in char_to_idx], dtype=torch.long)
|
49 |
+
|
50 |
+
def decode_to_text(predictions, idx_to_char):
|
51 |
+
return [''.join(idx_to_char[i] for i in pred if i in idx_to_char) for pred in predictions]
|
52 |
|
53 |
# Load fine-tuned Whisper model
|
54 |
+
transcriber_whisper = pipeline("automatic-speech-recognition", model="OwLim/whisper-sundanese-finetune")
|
55 |
+
transcriber_wav2vec = pipeline("automatic-speech-recognition", model="indonesian-nlp/wav2vec2-indonesian-javanese-sundanese")
|
56 |
+
|
57 |
+
|
58 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
59 |
+
SAMPLE_RATE = 16_000
|
60 |
+
|
61 |
+
model_path = hf_hub_download(repo_id="Blebbyblub/javanese-conformer-asrV2", filename="pytorch_model.bin")
|
62 |
+
model = ASRConformerModel(input_dim=80, vocab_size=29).to(device)
|
63 |
+
model.load_state_dict(torch.load(model_path, map_location=device))
|
64 |
+
|
65 |
+
examples_audio = [
|
66 |
+
file for file in os.listdir("./") if file.endswith(".wav")
|
67 |
+
]
|
68 |
+
|
69 |
+
idx_to_char = {v: k for k, v in char_to_idx.items()}
|
70 |
|
71 |
+
def transcribe(audio, model_selection):
|
72 |
sr, waveform = audio
|
73 |
+
|
74 |
# Change into Mono Audio
|
75 |
if waveform.ndim > 1:
|
76 |
waveform = waveform.mean(axis=1)
|
|
|
79 |
waveform = waveform.astype(np.float32)
|
80 |
waveform /= np.max(np.abs(waveform))
|
81 |
|
82 |
+
if "Conformer" == model_selection :
|
83 |
+
mel_transform = torchaudio.transforms.MelSpectrogram(sample_rate=SAMPLE_RATE, n_mels=80)
|
84 |
+
|
85 |
+
waveform = torch.from_numpy(waveform).float()
|
86 |
+
if sr != SAMPLE_RATE:
|
87 |
+
waveform = torchaudio.transforms.Resample(sr, SAMPLE_RATE)(waveform)
|
88 |
+
|
89 |
+
waveform = waveform.unsqueeze(0)
|
90 |
+
mel = mel_transform(waveform).squeeze(0).transpose(0, 1) # [time, mel]
|
91 |
+
mel = mel.unsqueeze(0).to(device)
|
92 |
+
input_length = torch.tensor([mel.size(1)]).to(device)
|
93 |
+
|
94 |
+
model.eval()
|
95 |
+
with torch.no_grad():
|
96 |
+
output, output_lengths = model(mel, input_length)
|
97 |
+
log_probs = output.log_softmax(2).transpose(0, 1)
|
98 |
+
pred_ids = greedy_decode(log_probs)
|
99 |
+
pred_text = decode_to_text(pred_ids, idx_to_char)[0]
|
100 |
+
|
101 |
+
return pred_text
|
102 |
+
|
103 |
+
if "Wav2Vec" == model_selection :
|
104 |
+
selected_model = transcriber_wav2vec
|
105 |
+
elif "Whisper" == model_selection:
|
106 |
+
selected_model = transcriber_whisper
|
107 |
+
|
108 |
+
return selected_model({
|
109 |
"sampling_rate" : sr,
|
110 |
"raw" : waveform
|
111 |
})["text"]
|
|
|
115 |
|
116 |
# --- Tab 1: Transcribe ---
|
117 |
with gr.Blocks() as tab_transcribe:
|
118 |
+
model_selector = gr.Radio(choices=["Whisper", "Conformer", "Wav2Vec"], label="Choose Model", info="This will effect the model that you use for transcribing", )
|
119 |
with gr.Row():
|
120 |
with gr.Column(scale=1):
|
121 |
audio_input = gr.Audio(sources="microphone", label="Record Your Voice")
|
|
|
126 |
with gr.Column(scale=1):
|
127 |
output_text = gr.Textbox(label="Transcription", placeholder="Waiting for Input", lines=3)
|
128 |
|
129 |
+
gr.Examples(
|
130 |
+
examples=examples_audio, # List of audio file paths
|
131 |
+
inputs=audio_input,
|
132 |
+
label="Try with Example Audio"
|
133 |
+
)
|
134 |
+
|
135 |
+
subBtn.click(fn=transcribe, inputs=[audio_input, model_selector], outputs=output_text)
|
136 |
clrBtn.click(fn=clear, outputs=[audio_input, output_text])
|
137 |
|
138 |
# --- Tab 2: Penjelasan Model Fine-Tuned ---
|
|
|
149 |
Model yang telah kami fine tune merupakan hasil <b>fine-tuning dari Whisper dan Conformer</b> untuk mendukung bahasa lokal di Indonesia, khususnya bahasa Jawa dan Sunda.
|
150 |
Model dilatih menggunakan kombinasi dataset <b>OpenSLR</b> berikut:
|
151 |
<br>
|
152 |
+
<a href="https://openslr.org/35/" target="_blank" style="text-decoration:none;">
|
153 |
+
<b>SLR35</b> - Large Javanese ASR training data set
|
154 |
+
</a>
|
|
|
155 |
<br>
|
156 |
+
<a href="https://openslr.org/36/" target="_blank" style="text-decoration:none;">
|
157 |
+
<b>SLR36</b> - Large Sundanese ASR training data set
|
158 |
+
</a>
|
159 |
+
<br>
|
160 |
+
<a href="https://openslr.org/41/" target="_blank" style="text-decoration:none;">
|
161 |
+
<b>SLR41</b> - High quality TTS data for Javanese
|
162 |
+
</a>
|
163 |
<br>
|
164 |
+
<a href="https://openslr.org/44" target="_blank" style="text-decoration:none;">
|
165 |
+
<b>SLR44</b> - High quality TTS data for Sundanese.
|
|
|
166 |
</a>
|
167 |
|
168 |
<br>
|
|
|
270 |
[tab_transcribe, tab_background, tab_architecture, tab_results, tab_authors],
|
271 |
["Transcribe", "Latar Belakang", "Arsitektur", "Evaluasi", "Fine-Tuned By"],
|
272 |
theme=gr.themes.Soft(),
|
273 |
+
title="Multilingual ASR Model"
|
274 |
)
|
275 |
|
276 |
if __name__ == "__main__":
|
jawa.wav
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:28e88d2a129ae797fde52637b187fef218c30eddc891b8189eed8c0b40bf9dec
|
3 |
+
size 200812
|
requirements.txt
CHANGED
@@ -1,3 +1,6 @@
|
|
|
|
|
|
1 |
numpy
|
2 |
torchaudio
|
3 |
-
transformers
|
|
|
|
1 |
+
os
|
2 |
+
torch
|
3 |
numpy
|
4 |
torchaudio
|
5 |
+
transformers
|
6 |
+
huggingface_hub
|
sunda.wav
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:bbc472fcba6f3f5203a9ccde45219f1dac1242a829451f6e397c905c3774eeac
|
3 |
+
size 615864
|
test.py
ADDED
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
examples_audio = [
|
3 |
+
'data/'+ file for file in os.listdir("data")
|
4 |
+
]
|
5 |
+
print(examples_audio)
|