Spaces:
Runtime error
Runtime error
add app.py
Browse files
app.py
ADDED
@@ -0,0 +1,69 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# import torch
|
2 |
+
# import torchaudio
|
3 |
+
import numpy as np
|
4 |
+
from espnet2.bin.st_inference_streaming import Speech2TextStreaming
|
5 |
+
import gradio as gr
|
6 |
+
import soundfile as sf
|
7 |
+
import librosa
|
8 |
+
|
9 |
+
# Load your custom model
|
10 |
+
model = Speech2TextStreaming(
|
11 |
+
st_model_file="/data1/ankita/st1/exp/st_train_st_raw_en_de_bpe_de2000_sp/valid.acc.ave_10best.pth", # path to your model weights
|
12 |
+
st_train_config="/data1/ankita/st1/exp/st_train_st_raw_en_de_bpe_de2000_sp/config.yaml", # path to your config file
|
13 |
+
device="cuda",
|
14 |
+
minlenratio=0.1,
|
15 |
+
maxlenratio=0.7,
|
16 |
+
beam_size=1 # change to "cuda" if using GPU
|
17 |
+
)
|
18 |
+
|
19 |
+
|
20 |
+
|
21 |
+
silence_threshold = 0.01 # Adjust this threshold based on your audio levels
|
22 |
+
silence_duration = 1.0 # Duration of silence to detect (in seconds)
|
23 |
+
|
24 |
+
def is_silence(audio_chunk, sr, threshold=silence_threshold):
|
25 |
+
return np.mean(np.abs(audio_chunk)) < threshold
|
26 |
+
|
27 |
+
def transcribe(state, new_chunk):
|
28 |
+
stream, silence_time = state
|
29 |
+
if new_chunk is None:
|
30 |
+
return (None, None), ""
|
31 |
+
|
32 |
+
sr, y = new_chunk
|
33 |
+
y = y.astype(np.float32)
|
34 |
+
|
35 |
+
if sr != 16000:
|
36 |
+
y = librosa.resample(y=y, orig_sr=sr, target_sr=16000)
|
37 |
+
y /= np.max(np.abs(y))
|
38 |
+
|
39 |
+
if stream is not None:
|
40 |
+
stream = np.concatenate([stream, y])
|
41 |
+
else:
|
42 |
+
stream = y
|
43 |
+
model(np.zeros(stream.shape), is_final=True)
|
44 |
+
|
45 |
+
if is_silence(y, sr):
|
46 |
+
silence_time += len(y) / sr
|
47 |
+
else:
|
48 |
+
silence_time = 0
|
49 |
+
|
50 |
+
if silence_time >= silence_duration:
|
51 |
+
output = model(stream, is_final=True)
|
52 |
+
return (None, 0), output[0][0] if output else ""
|
53 |
+
else:
|
54 |
+
output = model(stream)
|
55 |
+
return (stream, silence_time), output[0][0] if output else ""
|
56 |
+
|
57 |
+
def clear_transcription():
|
58 |
+
return (None, 0), ""
|
59 |
+
|
60 |
+
with gr.Blocks() as demo:
|
61 |
+
state = gr.State((None, 0))
|
62 |
+
audio = gr.Audio(sources=["microphone"], type="numpy", streaming=True)
|
63 |
+
text = gr.Textbox()
|
64 |
+
clear_button = gr.Button("Clear")
|
65 |
+
|
66 |
+
audio.stream(transcribe, inputs=[state, audio], outputs=[state, text])
|
67 |
+
clear_button.click(clear_transcription, inputs=[], outputs=[state, text])
|
68 |
+
|
69 |
+
demo.launch()
|