Spaces:
Running
on
Zero
Running
on
Zero
Update whisper.py
#5
by
ssolito
- opened
- whisper.py +87 -176
whisper.py
CHANGED
@@ -1,35 +1,59 @@
|
|
1 |
-
from pyannote.audio import Pipeline
|
2 |
from pydub import AudioSegment
|
3 |
import os
|
4 |
-
from transformers import WhisperForConditionalGeneration, WhisperProcessor
|
5 |
import torchaudio
|
6 |
import torch
|
7 |
import re
|
8 |
from transformers import pipeline
|
|
|
9 |
import spaces
|
10 |
|
11 |
-
|
12 |
device = 0 if torch.cuda.is_available() else "cpu"
|
13 |
torch_dtype = torch.float32
|
14 |
|
15 |
-
|
16 |
-
|
17 |
-
|
|
|
18 |
BATCH_SIZE = 1
|
19 |
-
model = WhisperForConditionalGeneration.from_pretrained(MODEL_NAME, torch_dtype=torch_dtype).to(device)
|
20 |
-
processor = WhisperProcessor.from_pretrained(MODEL_NAME)
|
21 |
-
pipeline_vad = Pipeline.from_pretrained("./pyannote/config.yaml")
|
22 |
-
threshold = 10000
|
23 |
-
segments_dir = "."
|
24 |
|
25 |
pipe = pipeline(
|
26 |
task="automatic-speech-recognition",
|
27 |
-
model=
|
28 |
chunk_length_s=30,
|
29 |
device=device,
|
30 |
token=os.getenv("HF_TOKEN")
|
31 |
)
|
32 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
33 |
def post_process_transcription(transcription, max_repeats=2):
|
34 |
tokens = re.findall(r'\b\w+\'?\w*\b[.,!?]?', transcription)
|
35 |
|
@@ -56,151 +80,6 @@ def post_process_transcription(transcription, max_repeats=2):
|
|
56 |
return cleaned_transcription
|
57 |
|
58 |
|
59 |
-
def convert_forced_to_tokens(forced_decoder_ids):
|
60 |
-
forced_decoder_tokens = []
|
61 |
-
for i, (idx, token) in enumerate(forced_decoder_ids):
|
62 |
-
if token is not None:
|
63 |
-
forced_decoder_tokens.append([idx, processor.tokenizer.decode(token)])
|
64 |
-
else:
|
65 |
-
forced_decoder_tokens.append([idx, token])
|
66 |
-
return forced_decoder_tokens
|
67 |
-
|
68 |
-
def generate_1st_chunk(audio):
|
69 |
-
|
70 |
-
input_audio, sample_rate = torchaudio.load(audio)
|
71 |
-
input_audio = torchaudio.transforms.Resample(sample_rate, 16000)(input_audio)
|
72 |
-
|
73 |
-
input_speech = input_audio[0]
|
74 |
-
|
75 |
-
input_features = processor(input_speech,
|
76 |
-
sampling_rate=16_000,
|
77 |
-
return_tensors="pt", torch_dtype=torch_dtype).input_features.to(device)
|
78 |
-
|
79 |
-
forced_decoder_ids = []
|
80 |
-
forced_decoder_ids.append([1,50270]) #[1, '<|ca|>']
|
81 |
-
forced_decoder_ids.append([2,50262]) #[2, '<|es|>']
|
82 |
-
forced_decoder_ids.append([3,50360]) #[3, '<|transcribe|>']
|
83 |
-
|
84 |
-
forced_decoder_ids_modified = forced_decoder_ids
|
85 |
-
idx = processor.tokenizer.all_special_tokens.index("<|startofprev|>")
|
86 |
-
forced_bos_token_id = processor.tokenizer.all_special_ids[idx]
|
87 |
-
prompt = "Antes de 'digui'm', '112'. 112, digui'm. Hola, puc parlar en castellà? Sí, digui, diga. Sí, mire: a veces al abrir la puerta de mi piso tengo una persona ahí. Vale, avisamos a la Guàrdia Urbana, ¿de acuerdo? Vale, perfecto. Gracias. Gracias. Buen día."
|
88 |
-
prompt_tokens = processor.tokenizer(prompt, add_special_tokens=False).input_ids
|
89 |
-
|
90 |
-
# we need to force these tokens
|
91 |
-
forced_decoder_ids = []
|
92 |
-
for idx, token in enumerate(prompt_tokens):
|
93 |
-
# indexing starts from 1 for forced tokens (token at position 0 is the SOS token)
|
94 |
-
forced_decoder_ids.append([idx + 1, token])
|
95 |
-
|
96 |
-
# now we add the SOS token at the end
|
97 |
-
offset = len(forced_decoder_ids)
|
98 |
-
forced_decoder_ids.append([offset + 1, model.generation_config.decoder_start_token_id])
|
99 |
-
|
100 |
-
# now we need to append the rest of the prefix tokens (lang, task, timestamps)
|
101 |
-
offset = len(forced_decoder_ids)
|
102 |
-
for idx, token in forced_decoder_ids_modified:
|
103 |
-
forced_decoder_ids.append([idx + offset , token])
|
104 |
-
|
105 |
-
model.generation_config.forced_decoder_ids = forced_decoder_ids
|
106 |
-
|
107 |
-
pred_ids = model.generate(input_features,
|
108 |
-
return_timestamps=True,
|
109 |
-
max_new_tokens=128,
|
110 |
-
decoder_start_token_id=forced_bos_token_id)
|
111 |
-
#exclude prompt from output
|
112 |
-
forced_decoder_tokens = convert_forced_to_tokens(forced_decoder_ids)
|
113 |
-
output = processor.decode(pred_ids[0][len(forced_decoder_tokens) + 1:], skip_special_tokens=True)
|
114 |
-
|
115 |
-
return output[1:]
|
116 |
-
|
117 |
-
def generate_2nd_chuk(audio):
|
118 |
-
|
119 |
-
input_audio, sample_rate = torchaudio.load(audio)
|
120 |
-
input_audio = torchaudio.transforms.Resample(sample_rate, 16000)(input_audio)
|
121 |
-
|
122 |
-
input_speech = input_audio[0]
|
123 |
-
|
124 |
-
input_features = processor(input_speech,
|
125 |
-
sampling_rate=16_000,
|
126 |
-
return_tensors="pt", torch_dtype=torch_dtype).input_features.to(device)
|
127 |
-
forced_decoder_ids = []
|
128 |
-
|
129 |
-
forced_decoder_ids.append([1,50270]) #[1, '<|ca|>']
|
130 |
-
forced_decoder_ids.append([2,50262]) #[2, '<|es|>']
|
131 |
-
forced_decoder_ids.append([3,50360]) #[3, '<|transcribe|>']
|
132 |
-
|
133 |
-
forced_decoder_ids_modified = forced_decoder_ids
|
134 |
-
idx = processor.tokenizer.all_special_tokens.index("<|startofprev|>")
|
135 |
-
forced_bos_token_id = processor.tokenizer.all_special_ids[idx]
|
136 |
-
|
137 |
-
prompt = "112, digui'm. Hola, puc parlar en castellà? Sí, digui, diga. Sí, mire: a veces al abrir la puerta de mi piso tengo una persona ahí. Vale, avisamos a la Guàrdia Urbana, ¿de acuerdo? Vale, perfecto. Gracias. Gracias. Buen día."
|
138 |
-
prompt_tokens = processor.tokenizer(prompt, add_special_tokens=False).input_ids
|
139 |
-
|
140 |
-
# we need to force these tokens
|
141 |
-
forced_decoder_ids = []
|
142 |
-
for idx, token in enumerate(prompt_tokens):
|
143 |
-
# indexing starts from 1 for forced tokens (token at position 0 is the SOS token)
|
144 |
-
forced_decoder_ids.append([idx + 1, token])
|
145 |
-
|
146 |
-
# now we add the SOS token at the end
|
147 |
-
offset = len(forced_decoder_ids)
|
148 |
-
forced_decoder_ids.append([offset + 1, model.generation_config.decoder_start_token_id])
|
149 |
-
|
150 |
-
# now we need to append the rest of the prefix tokens (lang, task, timestamps)
|
151 |
-
offset = len(forced_decoder_ids)
|
152 |
-
for idx, token in forced_decoder_ids_modified:
|
153 |
-
forced_decoder_ids.append([idx + offset , token])
|
154 |
-
|
155 |
-
model.generation_config.forced_decoder_ids = forced_decoder_ids
|
156 |
-
|
157 |
-
pred_ids = model.generate(input_features,
|
158 |
-
return_timestamps=True,
|
159 |
-
max_new_tokens=128,
|
160 |
-
decoder_start_token_id=forced_bos_token_id)
|
161 |
-
#exclude prompt from output
|
162 |
-
forced_decoder_tokens = convert_forced_to_tokens(forced_decoder_ids)
|
163 |
-
output = processor.decode(pred_ids[0][len(forced_decoder_tokens) + 1:], skip_special_tokens=True)
|
164 |
-
|
165 |
-
return output[1:]
|
166 |
-
|
167 |
-
def processing_vad_threshold(audio, output_vad, threshold, max_duration, concatenated_segment):
|
168 |
-
|
169 |
-
transcription_audio = ""
|
170 |
-
is_first_chunk = True
|
171 |
-
for speech in output_vad.get_timeline().support():
|
172 |
-
start, end = speech.start, speech.end
|
173 |
-
segment_duration = (end - start) * 1000
|
174 |
-
segment_audio = audio[start * 1000:end * 1000]
|
175 |
-
|
176 |
-
if max_duration + segment_duration < threshold:
|
177 |
-
concatenated_segment += audio[start * 1000:end * 1000]
|
178 |
-
max_duration += segment_duration
|
179 |
-
|
180 |
-
else:
|
181 |
-
if len(concatenated_segment) > 0:
|
182 |
-
temp_segment_path = os.path.join(segments_dir, f"temp_segment.wav")
|
183 |
-
concatenated_segment.export(temp_segment_path, format="wav")
|
184 |
-
|
185 |
-
if is_first_chunk:
|
186 |
-
output = generate_1st_chunk(temp_segment_path)
|
187 |
-
is_first_chunk = False
|
188 |
-
else:
|
189 |
-
output = generate_2nd_chuk(temp_segment_path)
|
190 |
-
transcription_audio = transcription_audio + output
|
191 |
-
max_duration = segment_duration
|
192 |
-
concatenated_segment = segment_audio
|
193 |
-
|
194 |
-
# Process any remaining audio in the concatenated_segment
|
195 |
-
if len(concatenated_segment) > 0:
|
196 |
-
temp_segment_path = os.path.join(segments_dir, f"temp_segment.wav")
|
197 |
-
concatenated_segment.export(temp_segment_path, format="wav")
|
198 |
-
|
199 |
-
output = generate_2nd_chuk(temp_segment_path)
|
200 |
-
transcription_audio = transcription_audio + output
|
201 |
-
|
202 |
-
return(transcription_audio)
|
203 |
-
|
204 |
def format_audio(audio_path):
|
205 |
input_audio, sample_rate = torchaudio.load(audio_path)
|
206 |
|
@@ -212,34 +91,66 @@ def format_audio(audio_path):
|
|
212 |
input_audio = input_audio.squeeze().numpy()
|
213 |
return(input_audio)
|
214 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
215 |
|
216 |
def transcribe_pipeline(audio, task):
|
217 |
text = pipe(audio, batch_size=BATCH_SIZE, generate_kwargs={"task": task}, return_timestamps=True)["text"]
|
218 |
return text
|
219 |
|
220 |
-
def generate(audio_path,
|
221 |
-
audio = AudioSegment.from_wav(audio_path)
|
222 |
|
223 |
-
|
224 |
-
|
225 |
-
|
226 |
-
|
227 |
-
|
228 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
229 |
|
230 |
-
|
231 |
-
|
232 |
-
|
233 |
-
|
234 |
-
|
235 |
-
|
236 |
-
|
237 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
238 |
output = transcribe_pipeline(format_audio(audio_path), task)
|
239 |
|
240 |
-
clean_output = post_process_transcription(output)
|
241 |
-
|
242 |
if temp_mono_path and os.path.exists(temp_mono_path):
|
243 |
-
|
|
|
|
|
|
|
|
|
244 |
|
245 |
return clean_output
|
|
|
|
|
1 |
from pydub import AudioSegment
|
2 |
import os
|
3 |
+
from transformers import WhisperForConditionalGeneration, WhisperProcessor, WhisperTokenizer
|
4 |
import torchaudio
|
5 |
import torch
|
6 |
import re
|
7 |
from transformers import pipeline
|
8 |
+
from peft import PeftModel, PeftConfig
|
9 |
import spaces
|
10 |
|
|
|
11 |
device = 0 if torch.cuda.is_available() else "cpu"
|
12 |
torch_dtype = torch.float32
|
13 |
|
14 |
+
### Configuration
|
15 |
+
MODEL_NAME_V2 = "./whisper-large-v3-catalan"
|
16 |
+
MODEL_NAME_V1 = "projecte-aina/whisper-large-v3-tiny-caesar"
|
17 |
+
CHUNK_LENGTH = 30
|
18 |
BATCH_SIZE = 1
|
|
|
|
|
|
|
|
|
|
|
19 |
|
20 |
pipe = pipeline(
|
21 |
task="automatic-speech-recognition",
|
22 |
+
model=MODEL_NAME_V1,
|
23 |
chunk_length_s=30,
|
24 |
device=device,
|
25 |
token=os.getenv("HF_TOKEN")
|
26 |
)
|
27 |
|
28 |
+
|
29 |
+
peft_config = PeftConfig.from_pretrained(MODEL_NAME_V2)
|
30 |
+
model = WhisperForConditionalGeneration.from_pretrained(
|
31 |
+
peft_config.base_model_name_or_path,
|
32 |
+
device_map="auto"
|
33 |
+
)
|
34 |
+
|
35 |
+
task = "transcribe"
|
36 |
+
|
37 |
+
model = PeftModel.from_pretrained(model, MODEL_NAME_V2)
|
38 |
+
model.config.use_cache = True
|
39 |
+
|
40 |
+
tokenizer = WhisperTokenizer.from_pretrained(peft_config.base_model_name_or_path, task=task)
|
41 |
+
processor = WhisperProcessor.from_pretrained(peft_config.base_model_name_or_path, task=task)
|
42 |
+
feature_extractor = processor.feature_extractor
|
43 |
+
forced_decoder_ids = processor.get_decoder_prompt_ids(task=task)
|
44 |
+
|
45 |
+
asr_pipe = pipeline(
|
46 |
+
task="automatic-speech-recognition",
|
47 |
+
model=model,
|
48 |
+
tokenizer=tokenizer,
|
49 |
+
feature_extractor=feature_extractor,
|
50 |
+
chunk_length_s=30)
|
51 |
+
|
52 |
+
def asr(audio_path, task):
|
53 |
+
asr_result = asr_pipe(audio_path, batch_size=BATCH_SIZE, generate_kwargs={"task":task}, return_timestamps=True)
|
54 |
+
base_model = asr_pipe.model.base_model if hasattr(asr_pipe.model, "base_model") else asr_pipe.model
|
55 |
+
return asr_result
|
56 |
+
|
57 |
def post_process_transcription(transcription, max_repeats=2):
|
58 |
tokens = re.findall(r'\b\w+\'?\w*\b[.,!?]?', transcription)
|
59 |
|
|
|
80 |
return cleaned_transcription
|
81 |
|
82 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
83 |
def format_audio(audio_path):
|
84 |
input_audio, sample_rate = torchaudio.load(audio_path)
|
85 |
|
|
|
91 |
input_audio = input_audio.squeeze().numpy()
|
92 |
return(input_audio)
|
93 |
|
94 |
+
def split_stereo_channels(audio_path):
|
95 |
+
|
96 |
+
audio = AudioSegment.from_wav(audio_path)
|
97 |
+
|
98 |
+
channels = audio.split_to_mono()
|
99 |
+
if len(channels) != 2:
|
100 |
+
raise ValueError(f"Audio {audio_path} does not have 2 channels.")
|
101 |
+
|
102 |
+
channels[0].export(f"temp_mono_speaker1.wav", format="wav") # Right
|
103 |
+
channels[1].export(f"temp_mono_speaker2.wav", format="wav") # Left
|
104 |
|
105 |
def transcribe_pipeline(audio, task):
|
106 |
text = pipe(audio, batch_size=BATCH_SIZE, generate_kwargs={"task": task}, return_timestamps=True)["text"]
|
107 |
return text
|
108 |
|
109 |
+
def generate(audio_path, use_v2):
|
|
|
110 |
|
111 |
+
if use_v2:
|
112 |
+
split_stereo_channels(audio_path)
|
113 |
+
|
114 |
+
audio_id = os.path.splitext(os.path.basename(audio_path))[0]
|
115 |
+
|
116 |
+
left_channel_path = "temp_mono_speaker2.wav"
|
117 |
+
right_channel_path = "temp_mono_speaker1.wav"
|
118 |
+
|
119 |
+
left_audio = format_audio(left_channel_path)
|
120 |
+
right_audio = format_audio(right_channel_path)
|
121 |
+
|
122 |
+
left_result = asr(left_audio, task)
|
123 |
+
right_result = asr(right_audio, task)
|
124 |
+
|
125 |
+
left_segs = [(seg["timestamp"][0], seg["timestamp"][1], "Speaker 1", post_process_transcription(seg["text"])) for seg in left_result["chunks"]]
|
126 |
+
right_segs = [(seg["timestamp"][0], seg["timestamp"][1], "Speaker 2", post_process_transcription(seg["text"])) for seg in right_result["chunks"]]
|
127 |
+
|
128 |
+
merged_transcript = sorted(left_segs + right_segs, key=lambda x: x[0])
|
129 |
+
merged_text = " ".join([seg[3] for seg in merged_transcript])
|
130 |
|
131 |
+
output = ""
|
132 |
+
for start, end, speaker, text in merged_transcript:
|
133 |
+
output += f"[{start:.2f}s - {end:.2f}s] {speaker}: {text}\n"
|
134 |
+
|
135 |
+
else:
|
136 |
+
audio = AudioSegment.from_wav(audio_path)
|
137 |
+
temp_mono_path = None
|
138 |
+
|
139 |
+
if audio.channels != 1: #stereo2mono
|
140 |
+
audio = audio.set_channels(1)
|
141 |
+
temp_mono_path = "temp_mono.wav"
|
142 |
+
audio.export(temp_mono_path, format="wav")
|
143 |
+
audio_path = temp_mono_path
|
144 |
+
task = "transcribe"
|
145 |
output = transcribe_pipeline(format_audio(audio_path), task)
|
146 |
|
147 |
+
clean_output = post_process_transcription(output, max_repeats=1) #check
|
148 |
+
|
149 |
if temp_mono_path and os.path.exists(temp_mono_path):
|
150 |
+
os.remove(temp_mono_path)
|
151 |
+
|
152 |
+
for temp_file in ["temp_mono_speaker1.wav", "temp_mono_speaker2.wav"]:
|
153 |
+
if os.path.exists(temp_file):
|
154 |
+
os.remove(temp_file)
|
155 |
|
156 |
return clean_output
|