File size: 13,379 Bytes
a95e64c
 
 
 
18680df
a95e64c
 
 
 
 
635f416
18680df
 
a95e64c
 
 
 
18680df
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a95e64c
18680df
 
 
 
 
635f416
18680df
 
 
 
 
 
a95e64c
 
 
 
 
 
 
 
 
 
c506d75
 
 
a08d5e7
a95e64c
 
c506d75
 
 
 
 
 
 
 
 
 
694b487
 
c506d75
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
694b487
 
c506d75
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
694b487
c506d75
 
 
a08d5e7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c506d75
 
 
 
 
 
 
 
a95e64c
c506d75
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
160effa
c506d75
 
 
 
 
 
635f416
694b487
a95e64c
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
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
import soundfile as sf
import datetime
from pyctcdecode import BeamSearchDecoderCTC
import torch
import json
import os
import time
import gc
import gradio as gr
import librosa
from transformers import Wav2Vec2ForCTC, AutoProcessor
from huggingface_hub import hf_hub_download
from torchaudio.models.decoder import ctc_decoder
from numba import cuda

# load pretrained model
model = Wav2Vec2ForCTC.from_pretrained("facebook/mms-1b-all")
processor = AutoProcessor.from_pretrained("facebook/mms-1b-all")

lm_decoding_config = {}
lm_decoding_configfile = hf_hub_download(
    repo_id="facebook/mms-cclms",
    filename="decoding_config.json",
    subfolder="mms-1b-all",
)

with open(lm_decoding_configfile) as f:
    lm_decoding_config = json.loads(f.read())

# allow language model decoding for "eng"

decoding_config = lm_decoding_config["amh"]

lm_file = hf_hub_download(
    repo_id="facebook/mms-cclms",
    filename=decoding_config["lmfile"].rsplit("/", 1)[1],
    subfolder=decoding_config["lmfile"].rsplit("/", 1)[0],
)
token_file = hf_hub_download(
    repo_id="facebook/mms-cclms",
    filename=decoding_config["tokensfile"].rsplit("/", 1)[1],
    subfolder=decoding_config["tokensfile"].rsplit("/", 1)[0],
)

beam_search_decoder = ctc_decoder(
    lexicon="./vocab_correct_cleaned.txt",
    tokens=token_file,
    lm=lm_file,
    nbest=1,
    beam_size=400,
    beam_size_token=50,
    lm_weight=float(decoding_config["lmweight"]),
    word_score=float(decoding_config["wordscore"]),
    sil_score=float(decoding_config["silweight"]),
    blank_token="<s>",
)

#Define Functions

#convert time into .sbv format
def format_time(seconds):
    # Convert seconds to hh:mm:ss,ms format
    return str(datetime.timedelta(seconds=seconds)).replace('.', ',')

#Convert Video/Audio into 16K wav file
def preprocessAudio(audioFile):
    if isinstance(audioFile, str):  # If audioFile is a string (filepath)
        os.system(f"ffmpeg -y -i {audioFile} -ar 16000 ./audioToConvert.wav")
    else:  # If audioFile is an object with a name attribute
        os.system(f"ffmpeg -y -i {audioFile.name} -ar 16000 ./audioToConvert.wav")

#Transcribe!!!
def Transcribe(file):
    try:
        device = "cuda:0" if torch.cuda.is_available() else "cpu"
        start_time = time.time()
        model.load_adapter("amh")
        processor.tokenizer.set_target_lang("amh")

        preprocessAudio(file)
        block_size = 30
        batch_size = 1
        filename = os.path.split(os.path.splitext(file.name)[0])[1]
        print(filename)

        transcripts = []
        speech_segments = []

        stream = librosa.stream(
            "./audioToConvert.wav",
            block_length=block_size,
            frame_length=16000,
            hop_length=16000
        )

        model.to(device)
        print(f"Model loaded to {device}: Entering transcription phase")

        #Code for timestamping
        encoding_start = 0
        encoding_end = 0
        sbv_file = open(f"{filename}_subtitle.sbv", "w")
        transcription_file = open(f"{filename}_transcription.txt", "w")

        # Create an empty list to hold batches
        batch = []

        for speech_segment in stream:
            if len(speech_segment.shape) > 1:
                speech_segment = speech_segment[:,0] + speech_segment[:,1]

            # Add the current speech segment to the batch
            batch.append(speech_segment)

            # If the batch is full, process it
            if len(batch) == batch_size:
                # Concatenate all segments in the batch along the time axis
                input_values = processor(batch, sampling_rate=16_000, return_tensors="pt", padding=True)
                input_values = input_values.to(device)
                with torch.no_grad():
                    logits = model(**input_values).logits
                if len(logits.shape) == 1:
                    logits = logits.unsqueeze(0)
                beam_search_result = beam_search_decoder(logits.to("cpu"))

                # Transcribe each segment in the batch
                for i in range(batch_size):
                    transcription = " ".join(beam_search_result[i][0].words).strip()
                    transcripts.append(transcription)

                    encoding_end = encoding_start + block_size
                    formatted_start = format_time(encoding_start)
                    formatted_end = format_time(encoding_end)
                    sbv_file.write(f"{formatted_start},{formatted_end}\n")
                    sbv_file.write(f"{transcription}\n\n")

                    encoding_start = encoding_end

                # Freeing up memory
                del input_values
                del logits
                del transcription
                torch.cuda.empty_cache()
                gc.collect()

                # Clear the batch
                batch = []

        if batch:
                # Concatenate all segments in the batch along the time axis
                input_values = processor(batch, sampling_rate=16_000, return_tensors="pt", padding=True)
                input_values = input_values.to(device)
                with torch.no_grad():
                    logits = model(**input_values).logits
                if len(logits.shape) == 1:
                    logits = logits.unsqueeze(0)
                beam_search_result = beam_search_decoder(logits.to("cpu"))

                # Transcribe each segment in the batch
                for i in range(len(batch)):
                    transcription = " ".join(beam_search_result[i][0].words).strip()
                    print(transcription)
                    transcripts.append(transcription)

                    encoding_end = encoding_start + block_size
                    formatted_start = format_time(encoding_start)
                    formatted_end = format_time(encoding_end)
                    sbv_file.write(f"{formatted_start},{formatted_end}\n")
                    sbv_file.write(f"{transcription}\n\n")

                    encoding_start = encoding_end

                # Freeing up memory
                del input_values
                del logits
                del transcription
                torch.cuda.empty_cache()
                gc.collect()

        # Join all transcripts into a single transcript
        transcript = ' '.join(transcripts)
        transcription_file.write(f"{transcript}")
        sbv_file.close()
        transcription_file.close()

        end_time = time.time()
        print(f"The script ran for {end_time - start_time} seconds.")
        return([f"./{filename}_subtitle.sbv", f"./{filename}_transcription.txt"])
    except Exception as e:
        error_log = open("error_log.txt", "w")
        error_log.write(f"Exception occurred: {e}")
        error_log.close()


#Transcribe!!!
def TranscribeMic(file):
    try:
        device = "cuda:0" if torch.cuda.is_available() else "cpu"
        start_time = time.time()
        model.load_adapter("amh")
        processor.tokenizer.set_target_lang("amh")

        preprocessAudio(file)
        block_size = 30
        batch_size = 1

        transcripts = []
        speech_segments = []

        stream = librosa.stream(
            "./audioToConvert.wav",
            block_length=block_size,
            frame_length=16000,
            hop_length=16000
        )

        model.to(device)
        print(f"Model loaded to {device}: Entering transcription phase")

        #Code for timestamping
        encoding_start = 0
        encoding_end = 0
        sbv_file = open(f"microphone_subtitle.sbv", "w")
        transcription_file = open(f"microphone_transcription.txt", "w")

        # Create an empty list to hold batches
        batch = []

        for speech_segment in stream:
            if len(speech_segment.shape) > 1:
                speech_segment = speech_segment[:,0] + speech_segment[:,1]

            # Add the current speech segment to the batch
            batch.append(speech_segment)

            # If the batch is full, process it
            if len(batch) == batch_size:
                # Concatenate all segments in the batch along the time axis
                input_values = processor(batch, sampling_rate=16_000, return_tensors="pt", padding=True)
                input_values = input_values.to(device)
                with torch.no_grad():
                    logits = model(**input_values).logits
                if len(logits.shape) == 1:
                    logits = logits.unsqueeze(0)
                beam_search_result = beam_search_decoder(logits.to("cpu"))

                # Transcribe each segment in the batch
                for i in range(batch_size):
                    transcription = " ".join(beam_search_result[i][0].words).strip()
                    transcripts.append(transcription)

                    encoding_end = encoding_start + block_size
                    formatted_start = format_time(encoding_start)
                    formatted_end = format_time(encoding_end)
                    sbv_file.write(f"{formatted_start},{formatted_end}\n")
                    sbv_file.write(f"{transcription}\n\n")

                    encoding_start = encoding_end

                # Freeing up memory
                del input_values
                del logits
                del transcription
                torch.cuda.empty_cache()
                gc.collect()

                # Clear the batch
                batch = []

        if batch:
                # Concatenate all segments in the batch along the time axis
                input_values = processor(batch, sampling_rate=16_000, return_tensors="pt", padding=True)
                input_values = input_values.to(device)
                with torch.no_grad():
                    logits = model(**input_values).logits
                if len(logits.shape) == 1:
                    logits = logits.unsqueeze(0)
                beam_search_result = beam_search_decoder(logits.to("cpu"))

                # Transcribe each segment in the batch
                for i in range(len(batch)):
                    transcription = " ".join(beam_search_result[i][0].words).strip()
                    print(transcription)
                    transcripts.append(transcription)

                    encoding_end = encoding_start + block_size
                    formatted_start = format_time(encoding_start)
                    formatted_end = format_time(encoding_end)
                    sbv_file.write(f"{formatted_start},{formatted_end}\n")
                    sbv_file.write(f"{transcription}\n\n")

                    encoding_start = encoding_end

                # Freeing up memory
                del input_values
                del logits
                del transcription
                torch.cuda.empty_cache()
                gc.collect()

        # Join all transcripts into a single transcript
        transcript = ' '.join(transcripts)
        transcription_file.write(f"{transcript}")
        sbv_file.close()
        transcription_file.close()

        end_time = time.time()
        print(f"The script ran for {end_time - start_time} seconds.")
        return([f"./microphone_subtitle.sbv", f"./microphone_transcription.txt"])
    except Exception as e:
        error_log = open("error_log.txt", "w")
        error_log.write(f"Exception occurred: {e}")
        error_log.close()

demo = gr.Blocks()

with demo:
    gr.Markdown(
    """
    <div align="center">
    
    # Amharic Audio Transcription

    This application uses Meta MMS and an Amharic kenLM model to transcribe Amharic Audio files of arbitrary length into .sbv and .txt files. Upload an Amharic audio file and get your transcription!

    ### (Note: Transcription quality is quite low, you should review and edit transcriptions before making them publicly available)
    </div>
    """)
    with gr.Tabs():
        with gr.TabItem("From File"):
            with gr.Row():
                file_input = gr.File(label="Upload an audio file of Amharic Content")
                file_output = gr.Files(label="Download output files")
            file_button = gr.Button("Submit")
        with gr.TabItem("From Microphone"):
            with gr.Row():
                microphone_input = gr.Audio(type="filepath", source="microphone")
                microphone_output = gr.Files(label="Download output files")
            microphone_button = gr.Button("Submit")
            
    file_button.click(Transcribe, inputs=file_input, outputs=file_output)
    microphone_button.click(TranscribeMic, inputs=microphone_input, outputs=microphone_output)
        
# demo = gr.Interface(fn=Transcribe, inputs=[gr.File(label="Upload an audio file of Amharic content"), gr.Slider(0, 25, value=4, step=1, label="batch size", info="Approximately .5GB per batch")], 
#                     outputs=gr.File(label="Download .sbv transcription", file_count="multiple"),
#                     title="Amharic Audio Transcription",
#                     description="This application uses Meta MMS and an Amharic kenLM model to transcribe Amharic Audio files of arbitrary length into .sbv and .txt files. Upload an Amharic audio file and get your transcription! \n(Note: Transcription quality is quite low, you should review and edit transcriptions before making them publicly available)"
#                     )

demo.queue(concurrency_count=10)
demo.launch()