Spaces:
Sleeping
Sleeping
Add application file
Browse files- app.py +69 -0
- packages.txt +1 -0
- requirements.txt +6 -0
app.py
ADDED
@@ -0,0 +1,69 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import nltk
|
2 |
+
import librosa
|
3 |
+
import torch
|
4 |
+
import kenlm
|
5 |
+
import gradio as gr
|
6 |
+
from pyctcdecode import build_ctcdecoder
|
7 |
+
from transformers import Wav2Vec2Processor,Wav2Vec2ProcessorWithLM,Wav2Vec2ForCTC
|
8 |
+
|
9 |
+
nltk.download("punkt")
|
10 |
+
|
11 |
+
def return_processor_and_model(model_name):
|
12 |
+
return Wav2Vec2Processor.from_pretrained(model_name), Wav2Vec2ForCTC.from_pretrained(model_name)
|
13 |
+
|
14 |
+
def return_processor_and_modelWithLM(model_name):
|
15 |
+
return Wav2Vec2ProcessorWithLM.from_pretrained(model_name), Wav2Vec2ForCTC.from_pretrained(model_name)
|
16 |
+
|
17 |
+
def load_and_fix_data(input_file):
|
18 |
+
speech, sample_rate = librosa.load(input_file)
|
19 |
+
if len(speech.shape) > 1:
|
20 |
+
speech = speech[:,0] + speech[:,1]
|
21 |
+
if sample_rate !=16000:
|
22 |
+
speech = librosa.resample(speech, sample_rate,16000)
|
23 |
+
return speech
|
24 |
+
|
25 |
+
def fix_transcription_casing(input_sentence):
|
26 |
+
sentences = nltk.sent_tokenize(input_sentence)
|
27 |
+
return (' '.join([s.replace(s[0],s[0].capitalize(),1) for s in sentences]))
|
28 |
+
|
29 |
+
|
30 |
+
def predict_and_ctc_lm_decode(input_file, model_name):
|
31 |
+
processor, model = return_processor_and_modelWithLM(model_name)
|
32 |
+
speech = load_and_fix_data(input_file)
|
33 |
+
|
34 |
+
input_values = processor(speech, return_tensors="pt", sampling_rate=16000).input_values
|
35 |
+
logits = model(input_values).logits.cpu().detach().numpy()[0]
|
36 |
+
|
37 |
+
pred = processor.batch_decode(logits.numpy()).text
|
38 |
+
|
39 |
+
transcribed_text = fix_transcription_casing(pred[0].lower())
|
40 |
+
|
41 |
+
return transcribed_text
|
42 |
+
|
43 |
+
def predict_and_greedy_decode(input_file, model_name):
|
44 |
+
processor, model = return_processor_and_model(model_name)
|
45 |
+
speech = load_and_fix_data(input_file)
|
46 |
+
|
47 |
+
input_values = processor(speech, return_tensors="pt", sampling_rate=16000).input_values
|
48 |
+
logits = model(input_values).logits
|
49 |
+
|
50 |
+
predicted_ids = torch.argmax(logits, dim=-1)
|
51 |
+
pred = processor.batch_decode(predicted_ids)
|
52 |
+
|
53 |
+
transcribed_text = fix_transcription_casing(pred[0].lower())
|
54 |
+
|
55 |
+
return transcribed_text
|
56 |
+
|
57 |
+
def return_all_predictions(input_file, model_name):
|
58 |
+
return predict_and_ctc_lm_decode(input_file, model_name), predict_and_greedy_decode(input_file, model_name)
|
59 |
+
|
60 |
+
|
61 |
+
gr.Interface(return_all_predictions,
|
62 |
+
inputs = [gr.inputs.Audio(source="microphone", type="filepath", label="Record/ Drop audio"), gr.inputs.Dropdown(["jonatasgrosman/wav2vec2-large-xlsr-53-spanish", "jonatasgrosman/wav2vec2-large-xlsr-53-spanish"], label="Model Name")],
|
63 |
+
outputs = [gr.outputs.Textbox(label="Beam CTC decoding w/ LM"), gr.outputs.Textbox(label="Greedy decoding")],
|
64 |
+
title="ASR using Wav2Vec2 & pyctcdecode in spanish",
|
65 |
+
description = "Comparing greedy decoder with beam search CTC decoder, record/ drop your audio!",
|
66 |
+
layout = "horizontal",
|
67 |
+
examples = [["test1.wav", "jonatasgrosman/wav2vec2-large-xlsr-53-spanish"], ["test2.wav", "jonatasgrosman/wav2vec2-large-xlsr-53-spanish"]],
|
68 |
+
theme="huggingface",
|
69 |
+
enable_queue=True).launch()
|
packages.txt
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
libsndfile1
|
requirements.txt
ADDED
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
nltk
|
2 |
+
transformers
|
3 |
+
torch
|
4 |
+
librosa
|
5 |
+
pyctcdecode
|
6 |
+
pypi-kenlm
|