csukuangfj
commited on
Commit
·
f0a085b
1
Parent(s):
6781708
small fixes
Browse files
app.py
CHANGED
@@ -21,8 +21,12 @@
|
|
21 |
|
22 |
|
23 |
import logging
|
|
|
|
|
24 |
import gradio as gr
|
25 |
-
|
|
|
|
|
26 |
|
27 |
title = "# Next-gen Kaldi: Generate subtitles for videos"
|
28 |
|
@@ -70,6 +74,11 @@ def build_html_output(s: str, style: str = "result_item_success"):
|
|
70 |
"""
|
71 |
|
72 |
|
|
|
|
|
|
|
|
|
|
|
73 |
def process_uploaded_file(
|
74 |
language: str,
|
75 |
repo_id: str,
|
@@ -84,7 +93,12 @@ def process_uploaded_file(
|
|
84 |
|
85 |
logging.info(f"Processing uploaded file: {in_filename}")
|
86 |
|
87 |
-
|
|
|
|
|
|
|
|
|
|
|
88 |
|
89 |
|
90 |
demo = gr.Blocks(css=css)
|
@@ -118,6 +132,7 @@ with demo:
|
|
118 |
source="upload",
|
119 |
interactive=True,
|
120 |
label="Upload from disk",
|
|
|
121 |
)
|
122 |
upload_button = gr.Button("Submit for recognition")
|
123 |
uploaded_output = gr.Textbox(label="Recognized speech from uploaded file")
|
|
|
21 |
|
22 |
|
23 |
import logging
|
24 |
+
import os
|
25 |
+
|
26 |
import gradio as gr
|
27 |
+
|
28 |
+
from decode import decode
|
29 |
+
from model import get_pretrained_model, get_vad, language_to_models
|
30 |
|
31 |
title = "# Next-gen Kaldi: Generate subtitles for videos"
|
32 |
|
|
|
74 |
"""
|
75 |
|
76 |
|
77 |
+
def show_file_info(in_filename: str):
|
78 |
+
logging.info(f"Input file: {in_filename}")
|
79 |
+
_ = os.system(f"ffprob -hide_banner -i '{in_filename}'")
|
80 |
+
|
81 |
+
|
82 |
def process_uploaded_file(
|
83 |
language: str,
|
84 |
repo_id: str,
|
|
|
93 |
|
94 |
logging.info(f"Processing uploaded file: {in_filename}")
|
95 |
|
96 |
+
recognizer = get_pretrained_model(repo_id)
|
97 |
+
vad = get_vad()
|
98 |
+
|
99 |
+
result = decode(recognizer, vad, in_filename)
|
100 |
+
|
101 |
+
return result, build_html_output("ok", "result_item_success")
|
102 |
|
103 |
|
104 |
demo = gr.Blocks(css=css)
|
|
|
132 |
source="upload",
|
133 |
interactive=True,
|
134 |
label="Upload from disk",
|
135 |
+
show_share_button=True,
|
136 |
)
|
137 |
upload_button = gr.Button("Submit for recognition")
|
138 |
uploaded_output = gr.Textbox(label="Recognized speech from uploaded file")
|
decode.py
ADDED
@@ -0,0 +1,117 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright 2022-2023 Xiaomi Corp. (authors: Fangjun Kuang)
|
2 |
+
#
|
3 |
+
# See LICENSE for clarification regarding multiple authors
|
4 |
+
#
|
5 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
6 |
+
# you may not use this file except in compliance with the License.
|
7 |
+
# You may obtain a copy of the License at
|
8 |
+
#
|
9 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
10 |
+
#
|
11 |
+
# Unless required by applicable law or agreed to in writing, software
|
12 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
13 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
14 |
+
# See the License for the specific language governing permissions and
|
15 |
+
# limitations under the License.
|
16 |
+
|
17 |
+
import subprocess
|
18 |
+
from dataclasses import dataclass
|
19 |
+
from datetime import timedelta
|
20 |
+
import logging
|
21 |
+
|
22 |
+
import numpy as np
|
23 |
+
import sherpa_onnx
|
24 |
+
|
25 |
+
from model import sample_rate
|
26 |
+
|
27 |
+
|
28 |
+
@dataclass
|
29 |
+
class Segment:
|
30 |
+
start: float
|
31 |
+
duration: float
|
32 |
+
text: str = ""
|
33 |
+
|
34 |
+
@property
|
35 |
+
def end(self):
|
36 |
+
return self.start + self.duration
|
37 |
+
|
38 |
+
def __str__(self):
|
39 |
+
s = f"{timedelta(seconds=self.start)}"[:-3]
|
40 |
+
s += " --> "
|
41 |
+
s += f"{timedelta(seconds=self.end)}"[:-3]
|
42 |
+
s = s.replace(".", ",")
|
43 |
+
s += "\n"
|
44 |
+
s += self.text
|
45 |
+
return s
|
46 |
+
|
47 |
+
|
48 |
+
def decode(
|
49 |
+
recognizer: sherpa_onnx.OfflineRecognizer,
|
50 |
+
vad: sherpa_onnx.VoiceActivityDetector,
|
51 |
+
filename: str,
|
52 |
+
) -> str:
|
53 |
+
ffmpeg_cmd = [
|
54 |
+
"ffmpeg",
|
55 |
+
"-i",
|
56 |
+
filename,
|
57 |
+
"-f",
|
58 |
+
"s16le",
|
59 |
+
"-acodec",
|
60 |
+
"pcm_s16le",
|
61 |
+
"-ac",
|
62 |
+
"1",
|
63 |
+
"-ar",
|
64 |
+
str(sample_rate),
|
65 |
+
"-",
|
66 |
+
]
|
67 |
+
|
68 |
+
process = subprocess.Popen(
|
69 |
+
ffmpeg_cmd, stdout=subprocess.PIPE, stderr=subprocess.DEVNULL
|
70 |
+
)
|
71 |
+
|
72 |
+
frames_per_read = int(sample_rate * 100) # 100 second
|
73 |
+
|
74 |
+
window_size = 512
|
75 |
+
|
76 |
+
buffer = []
|
77 |
+
|
78 |
+
segment_list = []
|
79 |
+
|
80 |
+
logging.info("Started!")
|
81 |
+
|
82 |
+
while True:
|
83 |
+
# *2 because int16_t has two bytes
|
84 |
+
data = process.stdout.read(frames_per_read * 2)
|
85 |
+
if not data:
|
86 |
+
break
|
87 |
+
|
88 |
+
samples = np.frombuffer(data, dtype=np.int16)
|
89 |
+
samples = samples.astype(np.float32) / 32768
|
90 |
+
|
91 |
+
buffer = np.concatenate([buffer, samples])
|
92 |
+
while len(buffer) > window_size:
|
93 |
+
vad.accept_waveform(buffer[:window_size])
|
94 |
+
buffer = buffer[window_size:]
|
95 |
+
|
96 |
+
streams = []
|
97 |
+
segments = []
|
98 |
+
while not vad.empty():
|
99 |
+
segment = Segment(
|
100 |
+
start=vad.front.start / sample_rate,
|
101 |
+
duration=len(vad.front.samples) / sample_rate,
|
102 |
+
)
|
103 |
+
segments.append(segment)
|
104 |
+
|
105 |
+
stream = recognizer.create_stream()
|
106 |
+
stream.accept_waveform(sample_rate, vad.front.samples)
|
107 |
+
|
108 |
+
streams.append(stream)
|
109 |
+
|
110 |
+
vad.pop()
|
111 |
+
|
112 |
+
recognizer.decode_streams(streams)
|
113 |
+
for seg, stream in zip(segments, streams):
|
114 |
+
seg.text = stream.result.text
|
115 |
+
segment_list.append(seg)
|
116 |
+
|
117 |
+
return "\n\n".join(f"{i}\n{seg} " for i, seg in enumerate(segment_list, 1))
|
model.py
CHANGED
@@ -165,6 +165,40 @@ def _get_russian_pre_trained_model(repo_id: str) -> sherpa_onnx.OfflineRecognize
|
|
165 |
return recognizer
|
166 |
|
167 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
168 |
english_models = {
|
169 |
"whisper-tiny.en": _get_whisper_model,
|
170 |
"whisper-base.en": _get_whisper_model,
|
|
|
165 |
return recognizer
|
166 |
|
167 |
|
168 |
+
@lru_cache(maxsize=2)
|
169 |
+
def get_vad() -> sherpa_onnx.VoiceActivityDetector:
|
170 |
+
vad_model = _get_nn_model_filename(
|
171 |
+
repo_id="csukuangfj/vad",
|
172 |
+
filename="silero_vad.onnx",
|
173 |
+
subfolder=".",
|
174 |
+
)
|
175 |
+
|
176 |
+
config = sherpa_onnx.VadModelConfig()
|
177 |
+
config.silero_vad.model = vad_model
|
178 |
+
config.silero_vad.min_silence_duration = 0.15
|
179 |
+
config.silero_vad.min_speech_duration = 0.25
|
180 |
+
config.sample_rate = sample_rate
|
181 |
+
|
182 |
+
vad = sherpa_onnx.VoiceActivityDetector(
|
183 |
+
config,
|
184 |
+
buffer_size_in_seconds=180,
|
185 |
+
)
|
186 |
+
|
187 |
+
return vad
|
188 |
+
|
189 |
+
|
190 |
+
@lru_cache(maxsize=10)
|
191 |
+
def get_pretrained_model(repo_id: str) -> sherpa_onnx.OfflineRecognizer:
|
192 |
+
if repo_id in english_models:
|
193 |
+
return english_models[repo_id](repo_id)
|
194 |
+
elif repo_id in chinese_english_mixed_models:
|
195 |
+
return chinese_english_mixed_models[repo_id](repo_id)
|
196 |
+
elif repo_id in russian_models:
|
197 |
+
return russian_models[repo_id](repo_id)
|
198 |
+
else:
|
199 |
+
raise ValueError(f"Unsupported repo_id: {repo_id}")
|
200 |
+
|
201 |
+
|
202 |
english_models = {
|
203 |
"whisper-tiny.en": _get_whisper_model,
|
204 |
"whisper-base.en": _get_whisper_model,
|