Spaces:
Sleeping
Sleeping
File size: 7,382 Bytes
bd540a9 d4bbf90 8cd0fcd bd540a9 d4bbf90 8cd0fcd d4bbf90 8cd0fcd d4bbf90 8cd0fcd bd540a9 005945b bd540a9 8cd0fcd bd540a9 d4bbf90 bd540a9 8cd0fcd d4bbf90 bd540a9 005945b d4bbf90 bd540a9 d4bbf90 bd540a9 d4bbf90 bd540a9 d4bbf90 bd540a9 9cece8a bd540a9 9cece8a bd540a9 d4bbf90 bd540a9 d4bbf90 bd540a9 d4bbf90 bd540a9 d4bbf90 bd540a9 d4bbf90 bd540a9 d4bbf90 bd540a9 d4bbf90 bd540a9 d4bbf90 e54a92c d4bbf90 e54a92c d4bbf90 e54a92c d4bbf90 bd540a9 d4bbf90 bd540a9 d4bbf90 bd540a9 d4bbf90 bd540a9 d4bbf90 bd540a9 d4bbf90 bd540a9 d4bbf90 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 |
import sys
import time
import torch
import torchaudio
import librosa
import gradio as gr
from transformers import AutoModelForCTC, Wav2Vec2BertProcessor
# Config
model_name = "Yehor/w2v-bert-2.0-uk-v2"
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
torch_dtype = torch.float16
min_duration = 0.5
max_duration = 60
concurrency_limit = 1
use_torch_compile = False
# Load the model
asr_model = AutoModelForCTC.from_pretrained(model_name, torch_dtype=torch_dtype).to(device)
processor = Wav2Vec2BertProcessor.from_pretrained(model_name)
if use_torch_compile:
asr_model = torch.compile(asr_model)
# Elements
examples = [
"example_1.wav",
"example_2.wav",
"example_3.wav",
"example_4.wav",
"example_5.wav",
"example_6.wav",
]
examples_table = '''
| File | Text |
| ------------- | ------------- |
| `example_1.wav` | тема про яку не люблять говорити офіційні джерела у генштабі і міноборони це хімічна зброя окупанти вже тривалий час використовують хімічну зброю заборонену |
| `example_2.wav` | всіма конвенціями якщо спочатку це були гранати з дронів то тепер фіксують випадки застосування |
| `example_3.wav` | хімічних снарядів причому склад отруйної речовони різний а отже й наслідки для наших військових теж різні |
| `example_4.wav` | використовує на фронті все що має і хімічна зброя не нийняток тож з чим маємо справу розбиралася марія моганисян |
| `example_5.wav` | двох тисяч випадків застосування росіянами боєприпасів споряджених небезпечними хімічними речовинами |
| `example_6.wav` | на всі писані норми марія моганисян олександр моторний спецкор марафон єдині новини |
'''.strip()
# https://www.tablesgenerator.com/markdown_tables
authors_table = '''
## Authors
Follow them in social networks and **contact** if you need any help or have any questions:
| <img src="https://avatars.githubusercontent.com/u/7875085?v=4" width="100"> **Yehor Smoliakov** |
|-------------------------------------------------------------------------------------------------|
| https://t.me/smlkw in Telegram |
| https://x.com/yehor_smoliakov at X |
| https://github.com/egorsmkv at GitHub |
| https://huggingface.co/Yehor at Hugging Face |
| or use egorsmkv@gmail.com |
'''.strip()
description_head = f"""
# Speech-to-Text for Ukrainian v2
## Overview
This space uses https://huggingface.co/Yehor/w2v-bert-2.0-uk-v2 model to recognize audio files.
> For demo, audio duration **must not** exceed **{max_duration}** seconds.
""".strip()
description_foot = f"""
## Community
- Join our Discord server where we talk about AI/ML/DL: https://discord.gg/yVAjkBgmt4
- Join our Speech Recognition group in Telegram: https://t.me/speech_recognition_uk
## More
Check out other ASR models: https://github.com/egorsmkv/speech-recognition-uk
{authors_table}
""".strip()
transcription_value = """
Recognized text will appear here.
Choose **an example file** below the Recognize button, upload **your audio file**, or use **the microphone** to record something.
""".strip()
tech_env = f"""
#### Environment
- Python: {sys.version}
- Torch device: {device}
- Torch dtype: {torch_dtype}
- Use torch.compile: {use_torch_compile}
""".strip()
tech_libraries = f"""
#### Libraries
- PyTorch: {torch.__version__}
- Transformers: {torch.__version__}
- Librosa: {librosa.version.version}
- Gradio: {gr.__version__}
""".strip()
def inference(audio_path, progress=gr.Progress()):
if not audio_path:
raise gr.Error("Please upload an audio file.")
gr.Info("Starting recognition", duration=2)
progress(0, desc="Recognizing")
duration = librosa.get_duration(path=audio_path)
if duration < min_duration:
raise gr.Error(f"The duration of the file is less than {min_duration} seconds, it is {round(duration, 2)} seconds.")
if duration > max_duration:
raise gr.Error(f"The duration of the file exceeds {max_duration} seconds.")
paths = [
audio_path,
]
results = []
for path in progress.tqdm(paths, desc="Recognizing...", unit="file"):
t0 = time.time()
audio_duration = librosa.get_duration(path=path, sr=16_000)
audio_input, _ = librosa.load(path, mono=True, sr=16_000)
features = processor([audio_input], sampling_rate=16_000).input_features
features = torch.tensor(features).to(device)
if torch_dtype == torch.float16:
features = features.half()
with torch.inference_mode():
logits = asr_model(features).logits
predicted_ids = torch.argmax(logits, dim=-1)
predictions = processor.batch_decode(predicted_ids)
if not predictions:
predictions = '-'
elapsed_time = round(time.time() - t0, 2)
rtf = round(elapsed_time / audio_duration, 4)
audio_duration = round(audio_duration, 2)
results.append(
{
"path": path.split("/")[-1],
"transcription": "\n".join(predictions),
"audio_duration": audio_duration,
"rtf": rtf,
}
)
gr.Info("Finished!", duration=2)
result_texts = []
for result in results:
result_texts.append(f'**{result["path"]}**')
result_texts.append("\n\n")
result_texts.append(f'> {result["transcription"]}')
result_texts.append("\n\n")
result_texts.append(f'**Audio duration**: {result["audio_duration"]}')
result_texts.append("\n")
result_texts.append(f'**Real-Time Factor**: {result["rtf"]}')
return "\n".join(result_texts)
demo = gr.Blocks(
title="Speech-to-Text for Ukrainian",
analytics_enabled=False,
)
with demo:
gr.Markdown(description_head)
gr.Markdown("## Demo")
with gr.Row():
audio_file = gr.Audio(label="Audio file", type="filepath")
transcription = gr.Markdown(
label="Transcription",
value=transcription_value,
)
gr.Button("Recognize").click(
inference,
concurrency_limit=concurrency_limit,
inputs=audio_file,
outputs=transcription,
)
with gr.Row():
gr.Examples(label="Choose an example", inputs=audio_file, examples=examples)
gr.Markdown(examples_table)
gr.Markdown(description_foot)
gr.Markdown('### Gradio app uses the following technologies:')
with gr.Row():
gr.Markdown(tech_env)
gr.Markdown(tech_libraries)
if __name__ == "__main__":
demo.queue()
demo.launch()
|