Commit
·
26ab82d
1
Parent(s):
6780507
add app.py
Browse files
app.py
ADDED
@@ -0,0 +1,83 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from transformers import pipeline
|
3 |
+
import gradio as gr
|
4 |
+
|
5 |
+
MODEL_NAME = "JackismyShephard/whisper-tiny-finetuned-minds14"
|
6 |
+
BATCH_SIZE = 8
|
7 |
+
|
8 |
+
device = 0 if torch.cuda.is_available() else "cpu"
|
9 |
+
|
10 |
+
pipe = pipeline(
|
11 |
+
task="automatic-speech-recognition",
|
12 |
+
model=MODEL_NAME,
|
13 |
+
chunk_length_s=30,
|
14 |
+
device=device,
|
15 |
+
)
|
16 |
+
|
17 |
+
|
18 |
+
# Copied from https://github.com/openai/whisper/blob/c09a7ae299c4c34c5839a76380ae407e7d785914/whisper/utils.py#L50
|
19 |
+
def format_timestamp(
|
20 |
+
seconds: float, always_include_hours: bool = False, decimal_marker: str = "."
|
21 |
+
):
|
22 |
+
if seconds is not None:
|
23 |
+
milliseconds = round(seconds * 1000.0)
|
24 |
+
|
25 |
+
hours = milliseconds // 3_600_000
|
26 |
+
milliseconds -= hours * 3_600_000
|
27 |
+
|
28 |
+
minutes = milliseconds // 60_000
|
29 |
+
milliseconds -= minutes * 60_000
|
30 |
+
|
31 |
+
seconds = milliseconds // 1_000
|
32 |
+
milliseconds -= seconds * 1_000
|
33 |
+
|
34 |
+
hours_marker = f"{hours:02d}:" if always_include_hours or hours > 0 else ""
|
35 |
+
return f"{hours_marker}{minutes:02d}:{seconds:02d}{decimal_marker}{milliseconds:03d}"
|
36 |
+
else:
|
37 |
+
# we have a malformed timestamp so just return it as is
|
38 |
+
return seconds
|
39 |
+
|
40 |
+
|
41 |
+
def transcribe(file, task, return_timestamps):
|
42 |
+
outputs = pipe(
|
43 |
+
file,
|
44 |
+
batch_size=BATCH_SIZE,
|
45 |
+
generate_kwargs={
|
46 |
+
"task": task,
|
47 |
+
"language": "english",
|
48 |
+
},
|
49 |
+
return_timestamps=return_timestamps,
|
50 |
+
)
|
51 |
+
text = outputs["text"]
|
52 |
+
if return_timestamps:
|
53 |
+
timestamps = outputs["chunks"]
|
54 |
+
timestamps = [
|
55 |
+
f"[{format_timestamp(chunk['timestamp'][0])} -> {format_timestamp(chunk['timestamp'][1])}] {chunk['text']}"
|
56 |
+
for chunk in timestamps
|
57 |
+
]
|
58 |
+
text = "\n".join(str(feature) for feature in timestamps)
|
59 |
+
return text
|
60 |
+
|
61 |
+
|
62 |
+
demo = gr.Interface(
|
63 |
+
fn=transcribe,
|
64 |
+
inputs=[
|
65 |
+
gr.Audio(label="Audio", type="filepath"),
|
66 |
+
gr.inputs.Radio(
|
67 |
+
["transcribe", "translate"], label="Task", default="transcribe"
|
68 |
+
),
|
69 |
+
gr.inputs.Checkbox(default=False, label="Return timestamps"),
|
70 |
+
],
|
71 |
+
outputs=gr.Textbox(),
|
72 |
+
title="Automatic Speech Recognition",
|
73 |
+
description=(
|
74 |
+
"Transcribe or translate long-form microphone or audio inputs with the click of a button! Demo uses the"
|
75 |
+
f" checkpoint [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) and 🤗 Transformers to transcribe or translate audio files"
|
76 |
+
" of arbitrary length."
|
77 |
+
),
|
78 |
+
examples="./examples",
|
79 |
+
cache_examples=True,
|
80 |
+
allow_flagging="never",
|
81 |
+
)
|
82 |
+
|
83 |
+
demo.launch()
|