Spaces:
Runtime error
Runtime error
Commit
•
3bc51c8
1
Parent(s):
71b5a2b
Update app.py
Browse files
app.py
CHANGED
@@ -1,9 +1,9 @@
|
|
1 |
import gradio as gr
|
2 |
-
|
3 |
import torch
|
4 |
import torchaudio
|
5 |
from datasets import load_dataset
|
6 |
-
from transformers import
|
7 |
import pandas as pd
|
8 |
from sklearn.model_selection import train_test_split
|
9 |
|
@@ -13,41 +13,17 @@ from pydub import AudioSegment
|
|
13 |
|
14 |
|
15 |
|
16 |
-
# processor = Wav2Vec2Processor.from_pretrained("RikeshSilwal/wav2vec2-nepali")
|
17 |
-
# model = Wav2Vec2ForCTC.from_pretrained("RikeshSilwal/wav2vec2-nepali")
|
18 |
-
|
19 |
-
processor = Wav2Vec2Processor.from_pretrained("RikeshSilwal/wav2vec2-nepali-rikeshsilwal")
|
20 |
-
model = Wav2Vec2ForCTC.from_pretrained("RikeshSilwal/wav2vec2-nepali-rikeshsilwal")
|
21 |
-
|
22 |
from torchaudio.transforms import Resample
|
23 |
import numpy as np
|
24 |
|
25 |
|
26 |
|
27 |
-
|
28 |
-
|
29 |
-
# def transcribe_audio(audio_file):
|
30 |
-
# input_arr, sampling_rate =torchaudio.load(audio_file)
|
31 |
-
# resampler = Resample(orig_freq=sampling_rate, new_freq=16000)
|
32 |
-
# input_arr = resampler(input_arr).squeeze().numpy()
|
33 |
-
# sampling_rate = 16000
|
34 |
-
# inputs = processor(input_arr, sampling_rate=16_000, return_tensors="pt", padding=True)
|
35 |
-
|
36 |
-
# with torch.no_grad():
|
37 |
-
# logits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits
|
38 |
-
|
39 |
-
# predicted_ids = torch.argmax(logits, dim=-1)
|
40 |
-
|
41 |
-
# predicted_words= processor.batch_decode(predicted_ids)
|
42 |
-
|
43 |
-
# return predicted_words[0]
|
44 |
-
|
45 |
def transcribe_audio(audio_file):
|
46 |
audio = AudioSegment.from_wav(audio_file)
|
47 |
|
48 |
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
|
49 |
|
50 |
-
input_arr, sampling_rate =
|
51 |
# Create TorchGating instance
|
52 |
tg = TG(sr=sampling_rate, nonstationary=True).to(device)
|
53 |
try:
|
@@ -55,13 +31,39 @@ def transcribe_audio(audio_file):
|
|
55 |
except:
|
56 |
input_arr = input_arr
|
57 |
if sampling_rate != 16000:
|
58 |
-
|
59 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
60 |
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
65 |
|
66 |
audio_input = gr.inputs.Audio(source="upload", type="filepath")
|
67 |
|
@@ -71,4 +73,3 @@ iface = gr.Interface(fn=transcribe_audio, inputs=audio_input,
|
|
71 |
button")
|
72 |
|
73 |
iface.launch(inline=False)
|
74 |
-
|
|
|
1 |
import gradio as gr
|
2 |
+
import librosa
|
3 |
import torch
|
4 |
import torchaudio
|
5 |
from datasets import load_dataset
|
6 |
+
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
|
7 |
import pandas as pd
|
8 |
from sklearn.model_selection import train_test_split
|
9 |
|
|
|
13 |
|
14 |
|
15 |
|
|
|
|
|
|
|
|
|
|
|
|
|
16 |
from torchaudio.transforms import Resample
|
17 |
import numpy as np
|
18 |
|
19 |
|
20 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
21 |
def transcribe_audio(audio_file):
|
22 |
audio = AudioSegment.from_wav(audio_file)
|
23 |
|
24 |
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
|
25 |
|
26 |
+
input_arr, sampling_rate =librosa.load(audio_file)
|
27 |
# Create TorchGating instance
|
28 |
tg = TG(sr=sampling_rate, nonstationary=True).to(device)
|
29 |
try:
|
|
|
31 |
except:
|
32 |
input_arr = input_arr
|
33 |
if sampling_rate != 16000:
|
34 |
+
input_arr = librosa.resample(input_arr, orig_sr=sampling_rate, target_sr=16000)
|
35 |
+
|
36 |
+
|
37 |
+
MODEL_NAME = "rikeshsilwalekg/whisper-small-wer35-ekg"
|
38 |
+
|
39 |
+
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
|
40 |
+
|
41 |
+
model = AutoModelForSpeechSeq2Seq.from_pretrained(
|
42 |
+
MODEL_NAME, torch_dtype=torch_dtype, use_safetensors=True
|
43 |
+
)
|
44 |
+
model.to(device)
|
45 |
|
46 |
+
processor = AutoProcessor.from_pretrained(MODEL_NAME)
|
47 |
+
|
48 |
+
pipe = pipeline(
|
49 |
+
"automatic-speech-recognition",
|
50 |
+
model=model,
|
51 |
+
tokenizer=processor.tokenizer,
|
52 |
+
feature_extractor=processor.feature_extractor,
|
53 |
+
max_new_tokens=128,
|
54 |
+
chunk_length_s=30,
|
55 |
+
batch_size=16,
|
56 |
+
return_timestamps=False,
|
57 |
+
torch_dtype=torch_dtype,
|
58 |
+
device=device,
|
59 |
+
)
|
60 |
+
|
61 |
+
# return_timestamps=True for sentence level timestaps
|
62 |
+
# for word level timestamps return_timestamps="word"
|
63 |
+
prediction = pipe(input_arr)
|
64 |
+
|
65 |
+
|
66 |
+
prediction = prediction['text']
|
67 |
|
68 |
audio_input = gr.inputs.Audio(source="upload", type="filepath")
|
69 |
|
|
|
73 |
button")
|
74 |
|
75 |
iface.launch(inline=False)
|
|