Spaces:
Sleeping
Sleeping
Upload 4 files
Browse files- README (1).md +14 -0
- app (2).py +233 -0
- gitattributes (1) +39 -0
- requirements (1).txt +26 -0
README (1).md
ADDED
@@ -0,0 +1,14 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
title: Indic Asr
|
3 |
+
emoji: 🏆
|
4 |
+
colorFrom: gray
|
5 |
+
colorTo: pink
|
6 |
+
sdk: gradio
|
7 |
+
sdk_version: 5.20.1
|
8 |
+
app_file: app.py
|
9 |
+
pinned: false
|
10 |
+
license: cc-by-4.0
|
11 |
+
short_description: A speech recognition tool for Indic languages.
|
12 |
+
---
|
13 |
+
|
14 |
+
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
|
app (2).py
ADDED
@@ -0,0 +1,233 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from __future__ import annotations
|
2 |
+
|
3 |
+
import os
|
4 |
+
|
5 |
+
import gradio as gr
|
6 |
+
|
7 |
+
import torch
|
8 |
+
import torchaudio
|
9 |
+
|
10 |
+
import spaces
|
11 |
+
|
12 |
+
import nemo.collections.asr as nemo_asr
|
13 |
+
|
14 |
+
LANGUAGE_NAME_TO_CODE = {
|
15 |
+
"Assamese": "as",
|
16 |
+
"Bengali": "bn",
|
17 |
+
"Bodo": "br",
|
18 |
+
"Dogri": "doi",
|
19 |
+
"Gujarati": "gu",
|
20 |
+
"Hindi": "hi",
|
21 |
+
"Kannada": "kn",
|
22 |
+
"Kashmiri": "ks",
|
23 |
+
"Konkani": "kok",
|
24 |
+
"Maithili": "mai",
|
25 |
+
"Malayalam": "ml",
|
26 |
+
"Manipuri": "mni",
|
27 |
+
"Marathi": "mr",
|
28 |
+
"Nepali": "ne",
|
29 |
+
"Odia": "or",
|
30 |
+
"Punjabi": "pa",
|
31 |
+
"Sanskrit": "sa",
|
32 |
+
"Santali": "sat",
|
33 |
+
"Sindhi": "sd",
|
34 |
+
"Tamil": "ta",
|
35 |
+
"Telugu": "te",
|
36 |
+
"Urdu": "ur"
|
37 |
+
}
|
38 |
+
|
39 |
+
|
40 |
+
DESCRIPTION = """\
|
41 |
+
### **IndicConformer: Speech Recognition for Indian Languages** 🎙️➡️📜
|
42 |
+
|
43 |
+
This Gradio demo showcases **IndicConformer**, a speech recognition model for **22 Indian languages**. The model operates in two modes: **CTC (Connectionist Temporal Classification)** and **RNNT (Recurrent Neural Network Transducer)**, providing robust and accurate transcriptions across diverse linguistic and acoustic conditions.
|
44 |
+
|
45 |
+
#### **How to Use:**
|
46 |
+
1. **Upload or record** an audio clip in any supported Indian language.
|
47 |
+
2. Select the **mode** (CTC or RNNT) for transcription.
|
48 |
+
3. Click **"Transcribe"** to generate the corresponding text in the target language.
|
49 |
+
4. View or copy the output for further use.
|
50 |
+
|
51 |
+
🚀 Try it out and experience seamless speech recognition for Indian languages!
|
52 |
+
"""
|
53 |
+
|
54 |
+
hf_token = os.getenv("HF_TOKEN")
|
55 |
+
device = "cuda:0" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"
|
56 |
+
torch_dtype = torch.bfloat16 if device != "cpu" else torch.float32
|
57 |
+
model_name_or_path = "ai4bharat/IndicConformer"
|
58 |
+
model = nemo_asr.models.EncDecCTCModel.from_pretrained(model_name_or_path).to(device)
|
59 |
+
# model = nemo_asr.models.EncDecCTCModel.restore_from("indicconformer_stt_bn_hybrid_rnnt_large.nemo").to(device)
|
60 |
+
model.eval()
|
61 |
+
|
62 |
+
CACHE_EXAMPLES = os.getenv("CACHE_EXAMPLES") == "1" and torch.cuda.is_available()
|
63 |
+
|
64 |
+
AUDIO_SAMPLE_RATE = 16000
|
65 |
+
MAX_INPUT_AUDIO_LENGTH = 60 # in seconds
|
66 |
+
DEFAULT_TARGET_LANGUAGE = "Bengali"
|
67 |
+
|
68 |
+
@spaces.GPU
|
69 |
+
def run_asr_ctc(input_audio: str, target_language: str) -> str:
|
70 |
+
lang_id = LANGUAGE_NAME_TO_CODE[target_language]
|
71 |
+
|
72 |
+
# Load and preprocess audio
|
73 |
+
audio_tensor, orig_freq = torchaudio.load(input_audio)
|
74 |
+
|
75 |
+
# Convert to mono if not already
|
76 |
+
if audio_tensor.shape[0] > 1:
|
77 |
+
audio_tensor = torch.mean(audio_tensor, dim=0, keepdim=True)
|
78 |
+
|
79 |
+
# Ensure shape [B x T]
|
80 |
+
if len(audio_tensor.shape) == 1:
|
81 |
+
audio_tensor = audio_tensor.unsqueeze(0) # Add batch dimension if missing
|
82 |
+
|
83 |
+
if audio_tensor.ndim > 1:
|
84 |
+
audio_tensor = audio_tensor.squeeze(0)
|
85 |
+
|
86 |
+
# Resample to 16kHz
|
87 |
+
audio_tensor = torchaudio.functional.resample(audio_tensor, orig_freq=orig_freq, new_freq=16000)
|
88 |
+
|
89 |
+
model.cur_decoder = "ctc"
|
90 |
+
ctc_text = model.transcribe([audio_tensor.numpy()], batch_size=1, logprobs=False, language_id=lang_id)[0]
|
91 |
+
|
92 |
+
return ctc_text[0]
|
93 |
+
|
94 |
+
# @spaces.GPU
|
95 |
+
# def run_asr_ctc(input_audio: str, target_language: str) -> str:
|
96 |
+
# # preprocess_audio(input_audio)
|
97 |
+
# # input_audio, orig_freq = torchaudio.load(input_audio)
|
98 |
+
# # input_audio = torchaudio.functional.resample(input_audio, orig_freq=orig_freq, new_freq=16000)
|
99 |
+
# lang_id = LANGUAGE_NAME_TO_CODE[target_language]
|
100 |
+
|
101 |
+
# model.cur_decoder = "ctc"
|
102 |
+
# ctc_text = model.transcribe([input_audio], batch_size=1, logprobs=False, language_id=lang_id)[0]
|
103 |
+
|
104 |
+
# return ctc_text[0]
|
105 |
+
|
106 |
+
@spaces.GPU
|
107 |
+
def run_asr_rnnt(input_audio: str, target_language: str) -> str:
|
108 |
+
lang_id = LANGUAGE_NAME_TO_CODE[target_language]
|
109 |
+
|
110 |
+
# Load and preprocess audio
|
111 |
+
audio_tensor, orig_freq = torchaudio.load(input_audio)
|
112 |
+
|
113 |
+
# Convert to mono if not already
|
114 |
+
if audio_tensor.shape[0] > 1:
|
115 |
+
audio_tensor = torch.mean(audio_tensor, dim=0, keepdim=True)
|
116 |
+
|
117 |
+
# Ensure shape [B x T]
|
118 |
+
if len(audio_tensor.shape) == 1:
|
119 |
+
audio_tensor = audio_tensor.unsqueeze(0) # Add batch dimension if missing
|
120 |
+
|
121 |
+
if audio_tensor.ndim > 1:
|
122 |
+
audio_tensor = audio_tensor.squeeze(0)
|
123 |
+
|
124 |
+
# Resample to 16kHz
|
125 |
+
audio_tensor = torchaudio.functional.resample(audio_tensor, orig_freq=orig_freq, new_freq=16000)
|
126 |
+
|
127 |
+
model.cur_decoder = "rnnt"
|
128 |
+
ctc_text = model.transcribe([audio_tensor.numpy()], batch_size=1, logprobs=False, language_id=lang_id)[0]
|
129 |
+
|
130 |
+
return ctc_text[0]
|
131 |
+
|
132 |
+
# @spaces.GPU
|
133 |
+
# def run_asr_rnnt(input_audio: str, target_language: str) -> str:
|
134 |
+
# # preprocess_audio(input_audio)
|
135 |
+
# # input_audio, orig_freq = torchaudio.load(input_audio)
|
136 |
+
# # input_audio = torchaudio.functional.resample(input_audio, orig_freq=orig_freq, new_freq=16000)
|
137 |
+
# lang_id = LANGUAGE_NAME_TO_CODE[target_language]
|
138 |
+
|
139 |
+
# model.cur_decoder = "rnnt"
|
140 |
+
# ctc_text = model.transcribe([input_audio], batch_size=1,logprobs=False, language_id=lang_id)[0]
|
141 |
+
|
142 |
+
# return ctc_text[0]
|
143 |
+
|
144 |
+
|
145 |
+
|
146 |
+
with gr.Blocks() as demo_asr_ctc:
|
147 |
+
with gr.Row():
|
148 |
+
with gr.Column():
|
149 |
+
with gr.Group():
|
150 |
+
input_audio = gr.Audio(label="Input speech", type="filepath")
|
151 |
+
target_language = gr.Dropdown(
|
152 |
+
label="Target language",
|
153 |
+
choices=LANGUAGE_NAME_TO_CODE.keys(),
|
154 |
+
value=DEFAULT_TARGET_LANGUAGE,
|
155 |
+
)
|
156 |
+
btn = gr.Button("Transcribe")
|
157 |
+
with gr.Column():
|
158 |
+
output_text = gr.Textbox(label="Transcribed text")
|
159 |
+
|
160 |
+
gr.Examples(
|
161 |
+
examples=[
|
162 |
+
["assets/Bengali.wav", "Bengali", "English"],
|
163 |
+
["assets/Gujarati.wav", "Gujarati", "Hindi"],
|
164 |
+
["assets/Punjabi.wav", "Punjabi", "Hindi"],
|
165 |
+
|
166 |
+
],
|
167 |
+
inputs=[input_audio, target_language],
|
168 |
+
outputs=output_text,
|
169 |
+
fn=run_asr_ctc,
|
170 |
+
cache_examples=CACHE_EXAMPLES,
|
171 |
+
api_name=False,
|
172 |
+
)
|
173 |
+
|
174 |
+
btn.click(
|
175 |
+
fn=run_asr_ctc,
|
176 |
+
inputs=[input_audio, target_language],
|
177 |
+
outputs=output_text,
|
178 |
+
api_name="asr",
|
179 |
+
)
|
180 |
+
|
181 |
+
with gr.Blocks() as demo_asr_rnnt:
|
182 |
+
with gr.Row():
|
183 |
+
with gr.Column():
|
184 |
+
with gr.Group():
|
185 |
+
input_audio = gr.Audio(label="Input speech", type="filepath")
|
186 |
+
target_language = gr.Dropdown(
|
187 |
+
label="Target language",
|
188 |
+
choices=LANGUAGE_NAME_TO_CODE.keys(),
|
189 |
+
value=DEFAULT_TARGET_LANGUAGE,
|
190 |
+
)
|
191 |
+
btn = gr.Button("Transcribe")
|
192 |
+
with gr.Column():
|
193 |
+
output_text = gr.Textbox(label="Transcribed text")
|
194 |
+
|
195 |
+
gr.Examples(
|
196 |
+
examples=[
|
197 |
+
["assets/Bengali.wav", "Bengali", "English"],
|
198 |
+
["assets/Gujarati.wav", "Gujarati", "Hindi"],
|
199 |
+
["assets/Punjabi.wav", "Punjabi", "Hindi"],
|
200 |
+
|
201 |
+
],
|
202 |
+
inputs=[input_audio, target_language],
|
203 |
+
outputs=output_text,
|
204 |
+
fn=run_asr_rnnt,
|
205 |
+
cache_examples=CACHE_EXAMPLES,
|
206 |
+
api_name=False,
|
207 |
+
)
|
208 |
+
|
209 |
+
btn.click(
|
210 |
+
fn=run_asr_rnnt,
|
211 |
+
inputs=[input_audio, target_language],
|
212 |
+
outputs=output_text,
|
213 |
+
api_name="asr",
|
214 |
+
)
|
215 |
+
|
216 |
+
|
217 |
+
with gr.Blocks(css="style.css") as demo:
|
218 |
+
gr.Markdown(DESCRIPTION)
|
219 |
+
gr.DuplicateButton(
|
220 |
+
value="Duplicate Space for private use",
|
221 |
+
elem_id="duplicate-button",
|
222 |
+
visible=os.getenv("SHOW_DUPLICATE_BUTTON") == "1",
|
223 |
+
)
|
224 |
+
|
225 |
+
with gr.Tabs():
|
226 |
+
with gr.Tab(label="CTC"):
|
227 |
+
demo_asr_ctc.render()
|
228 |
+
with gr.Tab(label="RNNT"):
|
229 |
+
demo_asr_rnnt.render()
|
230 |
+
|
231 |
+
|
232 |
+
if __name__ == "__main__":
|
233 |
+
demo.queue(max_size=50).launch()
|
gitattributes (1)
ADDED
@@ -0,0 +1,39 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
*.7z filter=lfs diff=lfs merge=lfs -text
|
2 |
+
*.arrow filter=lfs diff=lfs merge=lfs -text
|
3 |
+
*.bin filter=lfs diff=lfs merge=lfs -text
|
4 |
+
*.bz2 filter=lfs diff=lfs merge=lfs -text
|
5 |
+
*.ckpt filter=lfs diff=lfs merge=lfs -text
|
6 |
+
*.ftz filter=lfs diff=lfs merge=lfs -text
|
7 |
+
*.gz filter=lfs diff=lfs merge=lfs -text
|
8 |
+
*.h5 filter=lfs diff=lfs merge=lfs -text
|
9 |
+
*.joblib filter=lfs diff=lfs merge=lfs -text
|
10 |
+
*.lfs.* filter=lfs diff=lfs merge=lfs -text
|
11 |
+
*.mlmodel filter=lfs diff=lfs merge=lfs -text
|
12 |
+
*.model filter=lfs diff=lfs merge=lfs -text
|
13 |
+
*.msgpack filter=lfs diff=lfs merge=lfs -text
|
14 |
+
*.npy filter=lfs diff=lfs merge=lfs -text
|
15 |
+
*.npz filter=lfs diff=lfs merge=lfs -text
|
16 |
+
*.onnx filter=lfs diff=lfs merge=lfs -text
|
17 |
+
*.ot filter=lfs diff=lfs merge=lfs -text
|
18 |
+
*.parquet filter=lfs diff=lfs merge=lfs -text
|
19 |
+
*.pb filter=lfs diff=lfs merge=lfs -text
|
20 |
+
*.pickle filter=lfs diff=lfs merge=lfs -text
|
21 |
+
*.pkl filter=lfs diff=lfs merge=lfs -text
|
22 |
+
*.pt filter=lfs diff=lfs merge=lfs -text
|
23 |
+
*.pth filter=lfs diff=lfs merge=lfs -text
|
24 |
+
*.rar filter=lfs diff=lfs merge=lfs -text
|
25 |
+
*.safetensors filter=lfs diff=lfs merge=lfs -text
|
26 |
+
saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
27 |
+
*.tar.* filter=lfs diff=lfs merge=lfs -text
|
28 |
+
*.tar filter=lfs diff=lfs merge=lfs -text
|
29 |
+
*.tflite filter=lfs diff=lfs merge=lfs -text
|
30 |
+
*.tgz filter=lfs diff=lfs merge=lfs -text
|
31 |
+
*.wasm filter=lfs diff=lfs merge=lfs -text
|
32 |
+
*.xz 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 |
+
indicconformer_stt_bn_hybrid_rnnt_large.nemo filter=lfs diff=lfs merge=lfs -text
|
37 |
+
Bengali.wav filter=lfs diff=lfs merge=lfs -text
|
38 |
+
Gujarati.wav filter=lfs diff=lfs merge=lfs -text
|
39 |
+
Punjabi.wav filter=lfs diff=lfs merge=lfs -text
|
requirements (1).txt
ADDED
@@ -0,0 +1,26 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
git+https://github.com/AshwinSankar17/NeMo-ai4b@nemo-v2
|
2 |
+
torchaudio
|
3 |
+
pytorch-lightning==2.4.0
|
4 |
+
hydra-core==1.3.2
|
5 |
+
librosa==0.10.2.post1
|
6 |
+
sentencepiece==0.2.0
|
7 |
+
pandas==2.2.2
|
8 |
+
lhotse==1.27.0
|
9 |
+
editdistance==0.8.1
|
10 |
+
jiwer==3.0.4
|
11 |
+
pyannote.audio
|
12 |
+
webdataset==0.2.100
|
13 |
+
cython==0.29.37
|
14 |
+
pyyaml==6.0.2
|
15 |
+
argparse==1.4.0
|
16 |
+
onnxruntime==1.19.0
|
17 |
+
tqdm==4.66.5
|
18 |
+
transformers
|
19 |
+
huggingface_hub
|
20 |
+
tokenizers
|
21 |
+
datasets
|
22 |
+
inflect
|
23 |
+
IPython
|
24 |
+
soundfile
|
25 |
+
pydub
|
26 |
+
numpy<2.0
|