File size: 8,066 Bytes
0b03171
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
import gradio as gr
import json
import librosa
import os
import soundfile as sf
import tempfile
import uuid

from nemo.collections.asr.models import ASRModel

SAMPLE_RATE = 16000 # Hz

model = ASRModel.from_pretrained("nvidia/canary-1b")
model.eval()


MAX_AUDIO_SECONDS = 40


def convert_audio(audio_filepath, tmpdir, utt_id):
	"""
	Convert all files to monochannel 16 kHz wav files.
	Do not convert and raise error if audio too long.
	Returns output filename and duration.
	"""

	data, sr = librosa.load(audio_filepath)

	duration = librosa.get_duration(y=data, sr=sr)

	if duration > MAX_AUDIO_SECONDS:
		raise gr.Error(
			f"This demo can transcribe up to {MAX_AUDIO_SECONDS} seconds of audio."
		)

	if sr != SAMPLE_RATE:
		data = librosa.resample(data, orig_sr=sr, target_sr=SAMPLE_RATE)

	# monochannel
	data = librosa.to_mono(data)

	out_filename = os.path.join(tmpdir, utt_id + '.wav')

	# save output audio
	sf.write(out_filename, data, SAMPLE_RATE)

	return out_filename, duration


def transcribe(audio_filepath, src_lang, tgt_lang, pnc):

	if audio_filepath is None:
		raise gr.Error("Please provide some input audio: either upload an audio file or use the microphone")
	
	utt_id = uuid.uuid4()
	with tempfile.TemporaryDirectory() as tmpdir:

		converted_audio_filepath, duration = convert_audio(audio_filepath, tmpdir, str(utt_id))

		# map src_lang and tgt_lang from long versions to short
		LANG_LONG_TO_LANG_SHORT = {
			"English": "en",
			"Spanish": "es",
			"French": "fr",
			"German": "de",
		}
		if src_lang not in LANG_LONG_TO_LANG_SHORT.keys():
			raise ValueError(f"src_lang must be one of {LANG_LONG_TO_LANG_SHORT.keys()}")
		else:
			src_lang = LANG_LONG_TO_LANG_SHORT[src_lang]
		
		if tgt_lang not in LANG_LONG_TO_LANG_SHORT.keys():
			raise ValueError(f"tgt_lang must be one of {LANG_LONG_TO_LANG_SHORT.keys()}")
		else:
			tgt_lang = LANG_LONG_TO_LANG_SHORT[tgt_lang]
		

		# infer taskname from src_lang and tgt_lang
		if src_lang == tgt_lang:
			taskname = "asr"
		else:
			taskname = "s2t_translation"

		# update pnc variable to be "yes" or "no"
		pnc = "yes" if pnc else "no"

		# make manifest file and save
		manifest_data = {
			"audio_filepath": converted_audio_filepath,
			"source_lang": src_lang,
			"target_lang": tgt_lang,
			"taskname": taskname,
			"pnc": pnc,
			"answer": "predict",
			"duration": str(duration),
		}

		manifest_filepath = os.path.join(tmpdir, f'{utt_id}.json')

		with open(manifest_filepath, 'w') as fout:
			line = json.dumps(manifest_data)
			fout.write(line + '\n')

		# call transcribe, passing in manifest filepath
		model_output = model.transcribe(manifest_filepath)

	return model_output[0]

# add logic to make sure dropdown menus only suggest valid combos
def on_src_or_tgt_lang_change(src_lang_value, tgt_lang_value, pnc_value):
	"""Callback function for when src_lang or tgt_lang dropdown menus are changed.

	Args:
		src_lang_value(string), tgt_lang_value (string), pnc_value(bool) - the current 
			chosen "values" of each Gradio component
	Returns:
		src_lang, tgt_lang, pnc - these are the new Gradio components that will be displayed
	
	Note: I found the required logic is easier to understand if you think about the possible src & tgt langs as
	a matrix, e.g. with English, Spanish, French, German as the langs, and only transcription in the same language,
	and X -> English and English -> X translation being allowed, the matrix looks like the diagram below ("Y" means it is
	allowed to go into that state).
	It is easier to understand the code if you think about which state you are in, given the current src_lang_value and
	tgt_lang_value, and then which states you can go to from there.

			tgt lang
		- |EN |ES |FR |DE 
		------------------
		EN| Y | Y | Y | Y
		------------------
		src 	ES| Y | Y |   |
		lang	------------------
		FR| Y |   | Y |
		------------------
		DE| Y |   |   | Y
	"""

	if src_lang_value == "English" and tgt_lang_value == "English":
		# src_lang and tgt_lang can go anywhere
		src_lang = gr.Dropdown(
			choices=["English", "Spanish", "French", "German"],
			value=src_lang_value,
			label="Input audio is spoken in:"
		)
		tgt_lang = gr.Dropdown(
			choices=["English", "Spanish", "French", "German"],
			value=tgt_lang_value,
			label="Transcribe in language:"
		)
	elif src_lang_value == "English": 
		# src is English & tgt is non-English
		# => src can only be English or current tgt_lang_values
		# & tgt can be anything
		src_lang = gr.Dropdown(
			choices=["English", tgt_lang_value],
			value=src_lang_value,
			label="Input audio is spoken in:"
		)
		tgt_lang = gr.Dropdown(
			choices=["English", "Spanish", "French", "German"],
			value=tgt_lang_value,
			label="Transcribe in language:"
		)
	elif tgt_lang_value == "English": 
		# src is non-English & tgt is English
		# => src can be anything
		# & tgt can only be English or current src_lang_value
		src_lang = gr.Dropdown(
			choices=["English", "Spanish", "French", "German"],
			value=src_lang_value,
			label="Input audio is spoken in:"
		)
		tgt_lang = gr.Dropdown(
			choices=["English", src_lang_value],
			value=tgt_lang_value,
			label="Transcribe in language:"
		)
	else:
		# both src and tgt are non-English
		# => both src and tgt can only be switch to English or themselves
		src_lang = gr.Dropdown(
			choices=["English", src_lang_value],
			value=src_lang_value,
			label="Input audio is spoken in:"
		)
		tgt_lang = gr.Dropdown(
			choices=["English", tgt_lang_value],
			value=tgt_lang_value,
			label="Transcribe in language:"
		)
	# let pnc be anything if src_lang_value == tgt_lang_value, else fix to True
	if src_lang_value == tgt_lang_value:
		pnc = gr.Checkbox(
			value=pnc_value,
			label="Punctuation & Capitalization in transcript?",
			interactive=True
		)
	else:
		pnc = gr.Checkbox(
			value=True,
			label="Punctuation & Capitalization in transcript?",
			interactive=False
		)
	return src_lang, tgt_lang, pnc


with gr.Blocks(
	title="NeMo Canary Model",
	css="""
		textarea { font-size: 18px;}
		#model_output_text_box span {
			font-size: 18px;
			font-weight: bold;
		}
	""",
	theme=gr.themes.Default(text_size=gr.themes.sizes.text_lg) # make text slightly bigger (default is text_md )
) as demo:

	gr.HTML("<h1 style='text-align: center'>NeMo Canary model: Transcribe & Translate audio</h1>")

	with gr.Row():
		with gr.Column():
			gr.HTML("<p><b>Step 1:</b> Upload an audio file or record with your microphone.</p>")

			audio_file = gr.Audio(sources=["microphone", "upload"], type="filepath")

			gr.HTML("<p><b>Step 2:</b> Choose the input and output language.</p>")

			src_lang = gr.Dropdown(
				choices=["English", "Spanish", "French", "German"],
				value="English",
				label="Input audio is spoken in:"
			)

			with gr.Column():
				tgt_lang = gr.Dropdown(
					choices=["English", "Spanish", "French", "German"],
					value="English",
					label="Transcribe in language:"
				)
				pnc = gr.Checkbox(
					value=True,
					label="Punctuation & Capitalization in transcript?",
				)

		with gr.Column():

			gr.HTML("<p><b>Step 3:</b> Run the model.</p>")

			go_button = gr.Button(
				value="Run model",
				variant="primary", # make "primary" so it stands out (default is "secondary")
			)

			model_output_text_box = gr.Textbox(
				label="Model Output",
				elem_id="model_output_text_box",
			)

	with gr.Row():

		gr.HTML(
			"<p style='text-align: center'>"
				"🐤 <a href='#' target='_blank'>Canary model</a> | "
				"🧑‍💻 <a href='https://github.com/NVIDIA/NeMo' target='_blank'>NeMo Repository</a>"
			"</p>"
		)

	go_button.click(
		fn=transcribe, 
		inputs = [audio_file, src_lang, tgt_lang, pnc],
		outputs = [model_output_text_box]
	)

	# call on_src_or_tgt_lang_change whenever src_lang or tgt_lang dropdown menus are changed	
	src_lang.change(
		fn=on_src_or_tgt_lang_change,
		inputs=[src_lang, tgt_lang, pnc],
		outputs=[src_lang, tgt_lang, pnc],
	)
	tgt_lang.change(
		fn=on_src_or_tgt_lang_change,
		inputs=[src_lang, tgt_lang, pnc],
		outputs=[src_lang, tgt_lang, pnc],
	)


demo.queue()
demo.launch()