Spaces:
Sleeping
Sleeping
working demo
Browse files- app.py +23 -37
- requirements.txt +2 -0
app.py
CHANGED
@@ -1,64 +1,50 @@
|
|
1 |
import gradio as gr
|
2 |
import soundfile as sf
|
3 |
import torch
|
4 |
-
|
|
|
|
|
5 |
|
6 |
MODEL_NAME = "mikr/w2v-bert-2.0-czech-colab-cv16"
|
7 |
-
lang = "cs"
|
8 |
|
9 |
device = 0 if torch.cuda.is_available() else "cpu"
|
10 |
|
11 |
-
|
12 |
-
processor = Wav2Vec2Processor.from_pretrained(MODEL_NAME)
|
13 |
|
14 |
-
|
15 |
-
|
16 |
-
)
|
17 |
|
18 |
-
def transcribe(file_upload):
|
19 |
-
warn_output = ""
|
20 |
-
if (file_upload is None):
|
21 |
-
return "ERROR: You have to either use the microphone or upload an audio file"
|
22 |
|
23 |
-
|
24 |
-
|
25 |
-
return warn_output + text
|
26 |
|
|
|
|
|
27 |
|
28 |
-
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
-
if sr != 16000:
|
33 |
-
wlen = int(wav.shape[0] / sr * 16000)
|
34 |
-
wav = signal.resample(wav, wlen)
|
35 |
-
return wav
|
36 |
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
input_values = processor(wav, sampling_rate=16000).input_values[0]
|
41 |
-
input_values = torch.tensor(input_values, device=device).unsqueeze(0)
|
42 |
-
logits = model(input_values).logits
|
43 |
-
pred_ids = torch.argmax(logits, dim=-1)
|
44 |
-
xcp = processor.batch_decode(pred_ids)
|
45 |
-
return xcp[0]
|
46 |
|
47 |
|
48 |
iface = gr.Interface(
|
49 |
-
fn=
|
50 |
inputs=[
|
51 |
-
gr.File(type="
|
52 |
],
|
53 |
outputs="text",
|
54 |
theme="huggingface",
|
55 |
-
title="
|
56 |
description=(
|
57 |
-
"Transcribe
|
58 |
-
f" checkpoint [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) from
|
59 |
"and 🤗 Transformers to transcribe audio files of arbitrary length."
|
60 |
),
|
61 |
allow_flagging="never",
|
62 |
)
|
63 |
|
64 |
-
iface.launch()
|
|
|
1 |
import gradio as gr
|
2 |
import soundfile as sf
|
3 |
import torch
|
4 |
+
import numpy as np
|
5 |
+
import librosa
|
6 |
+
from transformers import AutoProcessor, Wav2Vec2BertForCTC
|
7 |
|
8 |
MODEL_NAME = "mikr/w2v-bert-2.0-czech-colab-cv16"
|
|
|
9 |
|
10 |
device = 0 if torch.cuda.is_available() else "cpu"
|
11 |
|
12 |
+
print("device:",device)
|
|
|
13 |
|
14 |
+
processor = AutoProcessor.from_pretrained(MODEL_NAME)
|
15 |
+
model = Wav2Vec2BertForCTC.from_pretrained(MODEL_NAME).to(device)
|
|
|
16 |
|
|
|
|
|
|
|
|
|
17 |
|
18 |
+
def transcribe(audio_path):
|
19 |
+
a, s = librosa.load(audio_path, sr=16_000)
|
|
|
20 |
|
21 |
+
# inputs = processor(a, sampling_rate=s, return_tensors="pt")
|
22 |
+
input_values = processor(a, sampling_rate=s, return_tensors="pt").input_features
|
23 |
|
24 |
+
with torch.no_grad():
|
25 |
+
logits = model(input_values.to(device)).logits
|
26 |
+
|
27 |
+
predicted_ids = torch.argmax(logits, dim=-1)
|
|
|
|
|
|
|
|
|
28 |
|
29 |
+
# transcribe speech
|
30 |
+
transcription = processor.batch_decode(predicted_ids)
|
31 |
+
return transcription[0]
|
|
|
|
|
|
|
|
|
|
|
|
|
32 |
|
33 |
|
34 |
iface = gr.Interface(
|
35 |
+
fn=transcribe,
|
36 |
inputs=[
|
37 |
+
gr.File(type="filepath", label="Upload Audio File"), # Audio file upload
|
38 |
],
|
39 |
outputs="text",
|
40 |
theme="huggingface",
|
41 |
+
title="Czech W2v-BERT 2.0 speech encoder demo - transcribe Czech Audio",
|
42 |
description=(
|
43 |
+
"Transcribe audio inputs with the click of a button! Demo uses the fine-tuned"
|
44 |
+
f" checkpoint [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) from Facebook W2v-BERT 2.0 speech encoder "
|
45 |
"and 🤗 Transformers to transcribe audio files of arbitrary length."
|
46 |
),
|
47 |
allow_flagging="never",
|
48 |
)
|
49 |
|
50 |
+
iface.launch(server_name="0.0.0.0")
|
requirements.txt
CHANGED
@@ -1,3 +1,5 @@
|
|
1 |
git+https://github.com/huggingface/transformers
|
2 |
torch
|
3 |
soundfile
|
|
|
|
|
|
1 |
git+https://github.com/huggingface/transformers
|
2 |
torch
|
3 |
soundfile
|
4 |
+
librosa
|
5 |
+
ffmpy
|