Spaces:
Sleeping
Sleeping
Initial commit
Browse files
gradio-app.py
ADDED
@@ -0,0 +1,193 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import subprocess
|
2 |
+
import time
|
3 |
+
|
4 |
+
import gradio as gr
|
5 |
+
import librosa
|
6 |
+
import pytube as pt
|
7 |
+
from models import asr, processor
|
8 |
+
from utils import format_timestamp
|
9 |
+
from vad import SpeechTimestampsMap, collect_chunks, get_speech_timestamps
|
10 |
+
|
11 |
+
## details: https://huggingface.co/docs/diffusers/optimization/fp16#automatic-mixed-precision-amp
|
12 |
+
# from torch import autocast
|
13 |
+
|
14 |
+
apply_vad = True
|
15 |
+
vad_parameters = {}
|
16 |
+
|
17 |
+
# task = "transcribe" # transcribe or translate
|
18 |
+
# language = "bn"
|
19 |
+
# asr.model.config.forced_decoder_ids = processor.get_decoder_prompt_ids(language=language, task=task)
|
20 |
+
# asr.model.config.max_new_tokens = 448 #default is 448
|
21 |
+
|
22 |
+
|
23 |
+
def _preprocess(filename):
|
24 |
+
audio_name = "audio.wav"
|
25 |
+
subprocess.call(
|
26 |
+
[
|
27 |
+
"ffmpeg",
|
28 |
+
"-y",
|
29 |
+
"-i",
|
30 |
+
filename,
|
31 |
+
"-acodec",
|
32 |
+
"pcm_s16le",
|
33 |
+
"-ar",
|
34 |
+
"16000",
|
35 |
+
"-ac",
|
36 |
+
"1",
|
37 |
+
"-loglevel",
|
38 |
+
"quiet",
|
39 |
+
audio_name,
|
40 |
+
]
|
41 |
+
)
|
42 |
+
return audio_name
|
43 |
+
|
44 |
+
|
45 |
+
def transcribe(microphone, file_upload):
|
46 |
+
warn_output = ""
|
47 |
+
if (microphone is not None) and (file_upload is not None):
|
48 |
+
warn_output = (
|
49 |
+
"WARNING: You've uploaded an audio file and used the microphone. "
|
50 |
+
"The recorded file from the microphone will be used and the uploaded audio will be discarded.\n"
|
51 |
+
)
|
52 |
+
|
53 |
+
elif (microphone is None) and (file_upload is None):
|
54 |
+
return "ERROR: You have to either use the microphone or upload an audio file"
|
55 |
+
|
56 |
+
file = microphone if microphone is not None else file_upload
|
57 |
+
print(f"\n\nFile is: {file}\n\n")
|
58 |
+
|
59 |
+
# for _preprocess(). No need if name of file provided in string format to asr pipeline as automatically uses ffmeg.
|
60 |
+
# Only required if ndarray given by using librosa.load() to load a file
|
61 |
+
start_time = time.time()
|
62 |
+
print("Starting Preprocessing")
|
63 |
+
# speech_array = _preprocess(filename=file)
|
64 |
+
filename = _preprocess(filename=file)
|
65 |
+
speech_array, sample_rate = librosa.load(f"{filename}", sr=16_000)
|
66 |
+
if apply_vad:
|
67 |
+
duration = speech_array.shape[0] / sample_rate
|
68 |
+
print(f"Processing audio with duration: {format_timestamp(duration)}")
|
69 |
+
speech_chunks = get_speech_timestamps(speech_array, **vad_parameters)
|
70 |
+
speech_array = collect_chunks(speech_array, speech_chunks)
|
71 |
+
print(f"VAD filter removed {format_timestamp(duration - (speech_array.shape[0] / sample_rate))}")
|
72 |
+
remaining_segments = ", ".join(
|
73 |
+
f'[{format_timestamp(chunk["start"] / sample_rate)} -> {format_timestamp(chunk["end"] / sample_rate)}]'
|
74 |
+
for chunk in speech_chunks
|
75 |
+
)
|
76 |
+
print(f"VAD filter kept the following audio segments: {remaining_segments}")
|
77 |
+
if not remaining_segments:
|
78 |
+
return "ERROR: No speech detected in the audio file"
|
79 |
+
|
80 |
+
|
81 |
+
|
82 |
+
print(f"\n Preprocessing COMPLETED in {round(time.time()-start_time, 2)}s \n")
|
83 |
+
|
84 |
+
start_time = time.time()
|
85 |
+
print("Starting Inference")
|
86 |
+
text = asr(speech_array)["text"]
|
87 |
+
# text = asr(file)["text"]
|
88 |
+
# with autocast("cuda"):
|
89 |
+
# text = asr(speech_array)["text"]
|
90 |
+
print(f"\n Inference COMPLETED in {round(time.time()-start_time, 2)}s \n")
|
91 |
+
|
92 |
+
return warn_output + text
|
93 |
+
|
94 |
+
|
95 |
+
def _return_yt_html_embed(yt_url):
|
96 |
+
if "?v=" in yt_url:
|
97 |
+
video_id = yt_url.split("?v=")[-1].split("&")[0]
|
98 |
+
else:
|
99 |
+
video_id = yt_url.split("/")[-1].split("?feature=")[0]
|
100 |
+
|
101 |
+
print(f"\n\nYT ID is: {video_id}\n\n")
|
102 |
+
return f'<center><iframe width="500" height="320" src="https://www.youtube.com/embed/{video_id}"> </iframe> </center>'
|
103 |
+
|
104 |
+
|
105 |
+
def yt_transcribe(yt_url):
|
106 |
+
start_time = time.time()
|
107 |
+
yt = pt.YouTube(yt_url)
|
108 |
+
html_embed_str = _return_yt_html_embed(yt_url)
|
109 |
+
stream = yt.streams.filter(only_audio=True)[0]
|
110 |
+
filename = "audio.mp3"
|
111 |
+
stream.download(filename=filename)
|
112 |
+
print(f"\n YT Audio Downloaded in {round(time.time()-start_time, 2)}s \n")
|
113 |
+
|
114 |
+
# for _preprocess(). No need if name of file provided in string format to asr pipeline as automatically uses ffmeg.
|
115 |
+
# Only required if ndarray given by using librosa.load() to load a file
|
116 |
+
start_time = time.time()
|
117 |
+
# print("Starting Preprocessing")
|
118 |
+
# speech_array = _preprocess(filename=filename)
|
119 |
+
# filename = _preprocess(filename=filename)
|
120 |
+
# speech_array, sample_rate = librosa.load(f"{filename}", sr=16_000)
|
121 |
+
# print(f"\n Preprocessing COMPLETED in {round(time.time()-start_time, 2)}s \n")
|
122 |
+
|
123 |
+
start_time = time.time()
|
124 |
+
print("Starting Inference")
|
125 |
+
text = asr(filename)["text"]
|
126 |
+
# with autocast("cuda"):
|
127 |
+
# text = asr(speech_array)["text"]
|
128 |
+
print(f"\n Inference COMPLETED in {round(time.time()-start_time, 2)}s \n")
|
129 |
+
|
130 |
+
return html_embed_str, text
|
131 |
+
|
132 |
+
|
133 |
+
mf_transcribe = gr.Interface(
|
134 |
+
fn=transcribe,
|
135 |
+
inputs=[
|
136 |
+
gr.Audio(source="microphone", type="filepath", label="Microphone"),
|
137 |
+
gr.Audio(source="upload", type="filepath", label="Upload File"),
|
138 |
+
],
|
139 |
+
outputs="text",
|
140 |
+
title="Bangla Demo: Transcribe Audio",
|
141 |
+
description=(
|
142 |
+
"Transcribe long-form microphone or audio inputs in BANGLA with the click of a button!"
|
143 |
+
),
|
144 |
+
allow_flagging="never",
|
145 |
+
)
|
146 |
+
|
147 |
+
yt_transcribe = gr.Interface(
|
148 |
+
fn=yt_transcribe,
|
149 |
+
inputs=[
|
150 |
+
gr.Textbox(
|
151 |
+
lines=1,
|
152 |
+
placeholder="Paste the URL to a Bangla language YouTube video here",
|
153 |
+
label="YouTube URL",
|
154 |
+
)
|
155 |
+
],
|
156 |
+
outputs=["html", "text"],
|
157 |
+
title="Bangla Demo: Transcribe YouTube",
|
158 |
+
description=(
|
159 |
+
"Transcribe long-form YouTube videos in BANGLA with the click of a button!"
|
160 |
+
),
|
161 |
+
allow_flagging="never",
|
162 |
+
)
|
163 |
+
# def transcribe2(audio, state=""):
|
164 |
+
# text = "text"
|
165 |
+
# state += text + " "
|
166 |
+
# return state, state
|
167 |
+
|
168 |
+
# Set the starting state to an empty string
|
169 |
+
|
170 |
+
# real_transcribe = gr.Interface(
|
171 |
+
# fn=transcribe2,
|
172 |
+
# inputs=[
|
173 |
+
# gr.Audio(source="microphone", type="filepath", streaming=True),
|
174 |
+
# "state"
|
175 |
+
# ],
|
176 |
+
# outputs=[
|
177 |
+
# "textbox",
|
178 |
+
# "state"
|
179 |
+
# ],
|
180 |
+
# live=True)
|
181 |
+
|
182 |
+
|
183 |
+
# demo = gr.TabbedInterface([mf_transcribe, yt_transcribe,real_transcribe], ["Transcribe Bangla Audio", "Transcribe Bangla YouTube Video","real time"])
|
184 |
+
demo = gr.TabbedInterface(
|
185 |
+
[mf_transcribe, yt_transcribe],
|
186 |
+
["Transcribe Bangla Audio", "Transcribe Bangla YouTube Video"],
|
187 |
+
)
|
188 |
+
|
189 |
+
|
190 |
+
if __name__ == "__main__":
|
191 |
+
demo.queue()
|
192 |
+
demo.launch(share="True")
|
193 |
+
# demo.launch(share='True', server_name="0.0.0.0", server_port=8080)
|
models.py
ADDED
@@ -0,0 +1,132 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
|
3 |
+
abs_path = os.path.abspath('.')
|
4 |
+
base_dir = os.path.dirname(os.path.dirname(abs_path))
|
5 |
+
os.environ['TRANSFORMERS_CACHE'] = os.path.join(base_dir, 'models_cache')
|
6 |
+
|
7 |
+
import torch
|
8 |
+
# Details: https://huggingface.co/docs/diffusers/optimization/fp16#enable-cudnn-autotuner
|
9 |
+
torch.backends.cudnn.benchmark = True
|
10 |
+
torch.backends.cuda.matmul.allow_tf32 = True
|
11 |
+
from transformers import pipeline, AutoTokenizer, AutoFeatureExtractor, AutoConfig, WhisperProcessor, WhisperForConditionalGeneration, WhisperTokenizer, WhisperFeatureExtractor
|
12 |
+
from typing import Union, BinaryIO
|
13 |
+
# from optimum.bettertransformer import BetterTransformer
|
14 |
+
|
15 |
+
language = '<|bn|>'
|
16 |
+
# language = '<|en|>'
|
17 |
+
task = "transcribe" # transcribe or translate
|
18 |
+
|
19 |
+
# model_name = 'openai/whisper-tiny.en'
|
20 |
+
# model_name = 'openai/whisper-base.en'
|
21 |
+
# model_name = 'openai/whisper-small.en'
|
22 |
+
# model_name = 'openai/whisper-medium'
|
23 |
+
## v2: trained on more epochs with regularization
|
24 |
+
# model_name = 'openai/whisper-large-v2'
|
25 |
+
|
26 |
+
## bangla
|
27 |
+
# model_name = 'Rakib/whisper-tiny-bn'
|
28 |
+
#model_name = 'anuragshas/whisper-small-bn'
|
29 |
+
# model_name = 'anuragshas/whisper-large-v2-bn'
|
30 |
+
# model_name = "Rakib/whisper-small-bn"
|
31 |
+
# model_name = "Rakib/whisper-small-bn-all"
|
32 |
+
# model_name = "Rakib/whisper-small-bn-all-600"
|
33 |
+
# model_name = "Rakib/whisper-small-bn-all-600-v2"
|
34 |
+
model_name = "Rakib/whisper-small-bn-crblp"
|
35 |
+
|
36 |
+
## lets you know the device count: cuda:0 or cuda:1
|
37 |
+
# print(torch.cuda.device_count())
|
38 |
+
|
39 |
+
device = 0 if torch.cuda.is_available() else -1
|
40 |
+
# device = -1 #Exclusively CPU
|
41 |
+
|
42 |
+
print(f"Using device: {'GPU' if device==0 else 'CPU'}")
|
43 |
+
|
44 |
+
if device !=0:
|
45 |
+
print("[Warning!] Using CPU could hamper performance")
|
46 |
+
|
47 |
+
print("Loading Tokenizer for ASR Speech-to-Text Model...\n" + "*" * 100)
|
48 |
+
# tokenizer = AutoTokenizer.from_pretrained(model_name, language=language, task=task)
|
49 |
+
# tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=True)
|
50 |
+
tokenizer = WhisperTokenizer.from_pretrained(model_name)
|
51 |
+
# tokenizer(['�', '�্র'],add_prefix_space=True, add_special_tokens=False).input_ids
|
52 |
+
|
53 |
+
print("Loading Feature Extractor for ASR Speech-to-Text Model...\n" + "*" * 100)
|
54 |
+
# feature_extractor = AutoFeatureExtractor.from_pretrained(model_name)
|
55 |
+
feature_extractor = WhisperFeatureExtractor.from_pretrained(model_name)
|
56 |
+
|
57 |
+
print("Loading Config for ASR Speech-to-Text Model...\n" + "*" * 100)
|
58 |
+
config = AutoConfig.from_pretrained(model_name)
|
59 |
+
|
60 |
+
print("Loading Processor for ASR Speech-to-Text Model...\n" + "*" * 100)
|
61 |
+
processor = WhisperProcessor(feature_extractor=feature_extractor, tokenizer=tokenizer)
|
62 |
+
|
63 |
+
print("Loading WHISPER ASR Speech-to-Text Model...\n" + "*" * 100)
|
64 |
+
model = WhisperForConditionalGeneration.from_pretrained(model_name)
|
65 |
+
|
66 |
+
## BetterTransformer (No Need if PyTorch 2.0 works!!)
|
67 |
+
## (currently 2secs faster inference than PyTorch 2.0 )
|
68 |
+
# model = WhisperForConditionalGeneration.from_pretrained(model_name)
|
69 |
+
# model = BetterTransformer.transform(model)
|
70 |
+
|
71 |
+
## bitsandbytes (only Linux & GPU) (requires conda env with conda-based pytorch!!!)
|
72 |
+
## currently only reduces size. slower inference than native models!!!
|
73 |
+
## from_pretrained doc: https://huggingface.co/docs/transformers/v4.25.1/en/main_classes/model#transformers.PreTrainedModel.from_pretrained
|
74 |
+
# model = WhisperForConditionalGeneration.from_pretrained(model_name, device_map="auto", load_in_8bit=True)
|
75 |
+
|
76 |
+
## For PyTorch 2.0 (Only Linux)
|
77 |
+
# model = WhisperForConditionalGeneration.from_pretrained(model_name).to(device="cuda:0")
|
78 |
+
##mode options are "default", "reduce-overhead" and "max-autotune". See: https://pytorch.org/get-started/pytorch-2.0/#modes
|
79 |
+
# model = torch.compile(model, mode="default")
|
80 |
+
|
81 |
+
|
82 |
+
asr = pipeline(
|
83 |
+
task="automatic-speech-recognition",
|
84 |
+
model=model,
|
85 |
+
tokenizer=tokenizer,
|
86 |
+
feature_extractor=feature_extractor,
|
87 |
+
# processor=processor, #no effect see: https://github.com/huggingface/transformers/blob/main/src/transformers/pipelines/automatic_speech_recognition.py
|
88 |
+
# config=config, #no effect see: https://github.com/huggingface/transformers/blob/main/src/transformers/pipelines/automatic_speech_recognition.py
|
89 |
+
device=device, # for gpu 1 for cpu -1
|
90 |
+
## chunk files longer than 30s into shorted samples
|
91 |
+
chunk_length_s=30,
|
92 |
+
## the amount of overlap (in secs) to be discarded while stitching the inferenced chunks
|
93 |
+
## stride_length_s is a tuple of the left and right stride(overlap) length.
|
94 |
+
## With only 1 number, both sides get the same stride, by default
|
95 |
+
## The stride_length on one side is 1/6th of the chunk_length_s if stride_length no provided
|
96 |
+
# stride_length_s=[8, 8],
|
97 |
+
stride_length_s=[5, 5],
|
98 |
+
# stride_length_s=[6,0],
|
99 |
+
batch_size=16,
|
100 |
+
ignore_warning=True,
|
101 |
+
## force whisper to generate timestamps so that the chunking and stitching can be accurate
|
102 |
+
# return_timestamps=True,
|
103 |
+
generate_kwargs = {
|
104 |
+
'language':language,
|
105 |
+
'task':task,
|
106 |
+
'repetition_penalty':1.8,
|
107 |
+
'num_beams':2,
|
108 |
+
'max_new_tokens':448,
|
109 |
+
'early_stopping':True,
|
110 |
+
# 'renormalize_logits':True,
|
111 |
+
# [16867]: �, [16867, 156, 100, 235, 156, 12811]: �্র
|
112 |
+
'bad_words_ids':[[16867], [16867, 156, 100, 235, 156, 12811]],
|
113 |
+
# 'supress_tokens': [16867, 156, 100, 235, 156, 12811],
|
114 |
+
}
|
115 |
+
)
|
116 |
+
|
117 |
+
|
118 |
+
def transcribe(speech_array: Union[str, BinaryIO], language: str = "en") -> str:
|
119 |
+
"""
|
120 |
+
Transcribes an audio array to text
|
121 |
+
Args:
|
122 |
+
speech_array (np.ndarray): audio in numpy array format
|
123 |
+
language (str): "sv" or "en"
|
124 |
+
Returns:
|
125 |
+
a string containing transcription
|
126 |
+
"""
|
127 |
+
asr.model.config.forced_decoder_ids = processor.get_decoder_prompt_ids(language=language, task=task)
|
128 |
+
# asr.model.config.max_new_tokens = 448 #default is 448
|
129 |
+
|
130 |
+
result = asr(speech_array)
|
131 |
+
|
132 |
+
return str(result["text"])
|
utils.py
ADDED
@@ -0,0 +1,23 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
|
2 |
+
|
3 |
+
def format_timestamp(
|
4 |
+
seconds: float,
|
5 |
+
always_include_hours: bool = False,
|
6 |
+
decimal_marker: str = ".",
|
7 |
+
) -> str:
|
8 |
+
assert seconds >= 0, "non-negative timestamp expected"
|
9 |
+
milliseconds = round(seconds * 1000.0)
|
10 |
+
|
11 |
+
hours = milliseconds // 3_600_000
|
12 |
+
milliseconds -= hours * 3_600_000
|
13 |
+
|
14 |
+
minutes = milliseconds // 60_000
|
15 |
+
milliseconds -= minutes * 60_000
|
16 |
+
|
17 |
+
seconds = milliseconds // 1_000
|
18 |
+
milliseconds -= seconds * 1_000
|
19 |
+
|
20 |
+
hours_marker = f"{hours:02d}:" if always_include_hours or hours > 0 else ""
|
21 |
+
return (
|
22 |
+
f"{hours_marker}{minutes:02d}:{seconds:02d}{decimal_marker}{milliseconds:03d}"
|
23 |
+
)
|
vad.py
ADDED
@@ -0,0 +1,273 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import bisect
|
2 |
+
import functools
|
3 |
+
import os
|
4 |
+
import warnings
|
5 |
+
|
6 |
+
from typing import List, Optional
|
7 |
+
|
8 |
+
import numpy as np
|
9 |
+
|
10 |
+
# The code below is adapted from https://github.com/snakers4/silero-vad.
|
11 |
+
|
12 |
+
def get_assets_path():
|
13 |
+
"""Returns the path to the assets directory."""
|
14 |
+
return os.path.join(os.path.dirname(os.path.abspath(__file__)), "assets")
|
15 |
+
|
16 |
+
|
17 |
+
def get_speech_timestamps(
|
18 |
+
audio: np.ndarray,
|
19 |
+
*,
|
20 |
+
threshold: float = 0.5,
|
21 |
+
# min_speech_duration_ms: int = 250,
|
22 |
+
min_speech_duration_ms: int = 800,
|
23 |
+
max_speech_duration_s: float = float("inf"),
|
24 |
+
# min_silence_duration_ms: int = 2000,
|
25 |
+
min_silence_duration_ms: int = 1000,
|
26 |
+
window_size_samples: int = 1024,
|
27 |
+
speech_pad_ms: int = 200,
|
28 |
+
) -> List[dict]:
|
29 |
+
"""This method is used for splitting long audios into speech chunks using silero VAD.
|
30 |
+
Args:
|
31 |
+
audio: One dimensional float array.
|
32 |
+
threshold: Speech threshold. Silero VAD outputs speech probabilities for each audio chunk,
|
33 |
+
probabilities ABOVE this value are considered as SPEECH. It is better to tune this
|
34 |
+
parameter for each dataset separately, but "lazy" 0.5 is pretty good for most datasets.
|
35 |
+
min_speech_duration_ms: Final speech chunks shorter min_speech_duration_ms are thrown out.
|
36 |
+
max_speech_duration_s: Maximum duration of speech chunks in seconds. Chunks longer
|
37 |
+
than max_speech_duration_s will be split at the timestamp of the last silence that
|
38 |
+
lasts more than 100s (if any), to prevent agressive cutting. Otherwise, they will be
|
39 |
+
split aggressively just before max_speech_duration_s.
|
40 |
+
min_silence_duration_ms: In the end of each speech chunk wait for min_silence_duration_ms
|
41 |
+
before separating it
|
42 |
+
window_size_samples: Audio chunks of window_size_samples size are fed to the silero VAD model.
|
43 |
+
WARNING! Silero VAD models were trained using 512, 1024, 1536 samples for 16000 sample rate.
|
44 |
+
Values other than these may affect model perfomance!!
|
45 |
+
speech_pad_ms: Final speech chunks are padded by speech_pad_ms each side
|
46 |
+
Returns:
|
47 |
+
List of dicts containing begin and end samples of each speech chunk.
|
48 |
+
"""
|
49 |
+
if window_size_samples not in [512, 1024, 1536]:
|
50 |
+
warnings.warn(
|
51 |
+
"Unusual window_size_samples! Supported window_size_samples:\n"
|
52 |
+
" - [512, 1024, 1536] for 16000 sampling_rate"
|
53 |
+
)
|
54 |
+
|
55 |
+
sampling_rate = 16000
|
56 |
+
min_speech_samples = sampling_rate * min_speech_duration_ms / 1000
|
57 |
+
speech_pad_samples = sampling_rate * speech_pad_ms / 1000
|
58 |
+
max_speech_samples = (
|
59 |
+
sampling_rate * max_speech_duration_s
|
60 |
+
- window_size_samples
|
61 |
+
- 2 * speech_pad_samples
|
62 |
+
)
|
63 |
+
min_silence_samples = sampling_rate * min_silence_duration_ms / 1000
|
64 |
+
min_silence_samples_at_max_speech = sampling_rate * 98 / 1000
|
65 |
+
|
66 |
+
audio_length_samples = len(audio)
|
67 |
+
|
68 |
+
model = get_vad_model()
|
69 |
+
state = model.get_initial_state(batch_size=1)
|
70 |
+
|
71 |
+
speech_probs = []
|
72 |
+
for current_start_sample in range(0, audio_length_samples, window_size_samples):
|
73 |
+
chunk = audio[current_start_sample : current_start_sample + window_size_samples]
|
74 |
+
if len(chunk) < window_size_samples:
|
75 |
+
chunk = np.pad(chunk, (0, int(window_size_samples - len(chunk))))
|
76 |
+
speech_prob, state = model(chunk, state, sampling_rate)
|
77 |
+
speech_probs.append(speech_prob)
|
78 |
+
|
79 |
+
triggered = False
|
80 |
+
speeches = []
|
81 |
+
current_speech = {}
|
82 |
+
neg_threshold = threshold - 0.15
|
83 |
+
|
84 |
+
# to save potential segment end (and tolerate some silence)
|
85 |
+
temp_end = 0
|
86 |
+
# to save potential segment limits in case of maximum segment size reached
|
87 |
+
prev_end = next_start = 0
|
88 |
+
|
89 |
+
for i, speech_prob in enumerate(speech_probs):
|
90 |
+
if (speech_prob >= threshold) and temp_end:
|
91 |
+
temp_end = 0
|
92 |
+
if next_start < prev_end:
|
93 |
+
next_start = window_size_samples * i
|
94 |
+
|
95 |
+
if (speech_prob >= threshold) and not triggered:
|
96 |
+
triggered = True
|
97 |
+
current_speech["start"] = window_size_samples * i
|
98 |
+
continue
|
99 |
+
|
100 |
+
if (
|
101 |
+
triggered
|
102 |
+
and (window_size_samples * i) - current_speech["start"] > max_speech_samples
|
103 |
+
):
|
104 |
+
if prev_end:
|
105 |
+
current_speech["end"] = prev_end
|
106 |
+
speeches.append(current_speech)
|
107 |
+
current_speech = {}
|
108 |
+
# previously reached silence (< neg_thres) and is still not speech (< thres)
|
109 |
+
if next_start < prev_end:
|
110 |
+
triggered = False
|
111 |
+
else:
|
112 |
+
current_speech["start"] = next_start
|
113 |
+
prev_end = next_start = temp_end = 0
|
114 |
+
else:
|
115 |
+
current_speech["end"] = window_size_samples * i
|
116 |
+
speeches.append(current_speech)
|
117 |
+
current_speech = {}
|
118 |
+
prev_end = next_start = temp_end = 0
|
119 |
+
triggered = False
|
120 |
+
continue
|
121 |
+
|
122 |
+
if (speech_prob < neg_threshold) and triggered:
|
123 |
+
if not temp_end:
|
124 |
+
temp_end = window_size_samples * i
|
125 |
+
# condition to avoid cutting in very short silence
|
126 |
+
if (window_size_samples * i) - temp_end > min_silence_samples_at_max_speech:
|
127 |
+
prev_end = temp_end
|
128 |
+
if (window_size_samples * i) - temp_end < min_silence_samples:
|
129 |
+
continue
|
130 |
+
else:
|
131 |
+
current_speech["end"] = temp_end
|
132 |
+
if (
|
133 |
+
current_speech["end"] - current_speech["start"]
|
134 |
+
) > min_speech_samples:
|
135 |
+
speeches.append(current_speech)
|
136 |
+
current_speech = {}
|
137 |
+
prev_end = next_start = temp_end = 0
|
138 |
+
triggered = False
|
139 |
+
continue
|
140 |
+
|
141 |
+
if (
|
142 |
+
current_speech
|
143 |
+
and (audio_length_samples - current_speech["start"]) > min_speech_samples
|
144 |
+
):
|
145 |
+
current_speech["end"] = audio_length_samples
|
146 |
+
speeches.append(current_speech)
|
147 |
+
|
148 |
+
for i, speech in enumerate(speeches):
|
149 |
+
if i == 0:
|
150 |
+
speech["start"] = int(max(0, speech["start"] - speech_pad_samples))
|
151 |
+
if i != len(speeches) - 1:
|
152 |
+
silence_duration = speeches[i + 1]["start"] - speech["end"]
|
153 |
+
if silence_duration < 2 * speech_pad_samples:
|
154 |
+
speech["end"] += int(silence_duration // 2)
|
155 |
+
speeches[i + 1]["start"] = int(
|
156 |
+
max(0, speeches[i + 1]["start"] - silence_duration // 2)
|
157 |
+
)
|
158 |
+
else:
|
159 |
+
speech["end"] = int(
|
160 |
+
min(audio_length_samples, speech["end"] + speech_pad_samples)
|
161 |
+
)
|
162 |
+
speeches[i + 1]["start"] = int(
|
163 |
+
max(0, speeches[i + 1]["start"] - speech_pad_samples)
|
164 |
+
)
|
165 |
+
else:
|
166 |
+
speech["end"] = int(
|
167 |
+
min(audio_length_samples, speech["end"] + speech_pad_samples)
|
168 |
+
)
|
169 |
+
|
170 |
+
return speeches
|
171 |
+
|
172 |
+
|
173 |
+
def collect_chunks(audio: np.ndarray, chunks: List[dict]) -> np.ndarray:
|
174 |
+
"""Collects and concatenates audio chunks."""
|
175 |
+
if not chunks:
|
176 |
+
return np.array([], dtype=np.float32)
|
177 |
+
|
178 |
+
return np.concatenate([audio[chunk["start"] : chunk["end"]] for chunk in chunks])
|
179 |
+
|
180 |
+
|
181 |
+
class SpeechTimestampsMap:
|
182 |
+
"""Helper class to restore original speech timestamps."""
|
183 |
+
|
184 |
+
def __init__(self, chunks: List[dict], sampling_rate: int, time_precision: int = 2):
|
185 |
+
self.sampling_rate = sampling_rate
|
186 |
+
self.time_precision = time_precision
|
187 |
+
self.chunk_end_sample = []
|
188 |
+
self.total_silence_before = []
|
189 |
+
|
190 |
+
previous_end = 0
|
191 |
+
silent_samples = 0
|
192 |
+
|
193 |
+
for chunk in chunks:
|
194 |
+
silent_samples += chunk["start"] - previous_end
|
195 |
+
previous_end = chunk["end"]
|
196 |
+
|
197 |
+
self.chunk_end_sample.append(chunk["end"] - silent_samples)
|
198 |
+
self.total_silence_before.append(silent_samples / sampling_rate)
|
199 |
+
|
200 |
+
def get_original_time(
|
201 |
+
self,
|
202 |
+
time: float,
|
203 |
+
chunk_index: Optional[int] = None,
|
204 |
+
) -> float:
|
205 |
+
if chunk_index is None:
|
206 |
+
chunk_index = self.get_chunk_index(time)
|
207 |
+
|
208 |
+
total_silence_before = self.total_silence_before[chunk_index]
|
209 |
+
return round(total_silence_before + time, self.time_precision)
|
210 |
+
|
211 |
+
def get_chunk_index(self, time: float) -> int:
|
212 |
+
sample = int(time * self.sampling_rate)
|
213 |
+
return min(
|
214 |
+
bisect.bisect(self.chunk_end_sample, sample),
|
215 |
+
len(self.chunk_end_sample) - 1,
|
216 |
+
)
|
217 |
+
|
218 |
+
|
219 |
+
@functools.lru_cache
|
220 |
+
def get_vad_model():
|
221 |
+
"""Returns the VAD model instance."""
|
222 |
+
path = os.path.join(get_assets_path(), "silero_vad.onnx")
|
223 |
+
return SileroVADModel(path)
|
224 |
+
|
225 |
+
|
226 |
+
class SileroVADModel:
|
227 |
+
def __init__(self, path):
|
228 |
+
try:
|
229 |
+
import onnxruntime
|
230 |
+
except ImportError as e:
|
231 |
+
raise RuntimeError(
|
232 |
+
"Applying the VAD filter requires the onnxruntime package"
|
233 |
+
) from e
|
234 |
+
|
235 |
+
opts = onnxruntime.SessionOptions()
|
236 |
+
opts.inter_op_num_threads = 1
|
237 |
+
opts.intra_op_num_threads = 1
|
238 |
+
opts.log_severity_level = 4
|
239 |
+
|
240 |
+
self.session = onnxruntime.InferenceSession(
|
241 |
+
path,
|
242 |
+
providers=["CPUExecutionProvider"],
|
243 |
+
sess_options=opts,
|
244 |
+
)
|
245 |
+
|
246 |
+
def get_initial_state(self, batch_size: int):
|
247 |
+
h = np.zeros((2, batch_size, 64), dtype=np.float32)
|
248 |
+
c = np.zeros((2, batch_size, 64), dtype=np.float32)
|
249 |
+
return h, c
|
250 |
+
|
251 |
+
def __call__(self, x, state, sr: int):
|
252 |
+
if len(x.shape) == 1:
|
253 |
+
x = np.expand_dims(x, 0)
|
254 |
+
if len(x.shape) > 2:
|
255 |
+
raise ValueError(
|
256 |
+
f"Too many dimensions for input audio chunk {len(x.shape)}"
|
257 |
+
)
|
258 |
+
if sr / x.shape[1] > 31.25:
|
259 |
+
raise ValueError("Input audio chunk is too short")
|
260 |
+
|
261 |
+
h, c = state
|
262 |
+
|
263 |
+
ort_inputs = {
|
264 |
+
"input": x,
|
265 |
+
"h": h,
|
266 |
+
"c": c,
|
267 |
+
"sr": np.array(sr, dtype="int64"),
|
268 |
+
}
|
269 |
+
|
270 |
+
out, h, c = self.session.run(None, ort_inputs)
|
271 |
+
state = (h, c)
|
272 |
+
|
273 |
+
return out, state
|