Spaces:
Running
on
T4
Running
on
T4
erastorgueva-nv
commited on
Commit
•
a28041c
1
Parent(s):
9379375
allow buffered inference up to 10 mins
Browse files- app.py +53 -18
- requirements.txt +1 -1
app.py
CHANGED
@@ -6,16 +6,39 @@ import soundfile as sf
|
|
6 |
import tempfile
|
7 |
import uuid
|
8 |
|
|
|
|
|
9 |
from nemo.collections.asr.models import ASRModel
|
|
|
|
|
10 |
|
11 |
SAMPLE_RATE = 16000 # Hz
|
|
|
12 |
|
13 |
model = ASRModel.from_pretrained("nvidia/canary-1b")
|
14 |
model.eval()
|
15 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
16 |
|
17 |
-
|
|
|
|
|
|
|
|
|
|
|
18 |
|
|
|
19 |
|
20 |
def convert_audio(audio_filepath, tmpdir, utt_id):
|
21 |
"""
|
@@ -24,21 +47,20 @@ def convert_audio(audio_filepath, tmpdir, utt_id):
|
|
24 |
Returns output filename and duration.
|
25 |
"""
|
26 |
|
27 |
-
data, sr = librosa.load(audio_filepath)
|
28 |
|
29 |
duration = librosa.get_duration(y=data, sr=sr)
|
30 |
|
31 |
-
if duration >
|
32 |
raise gr.Error(
|
33 |
-
f"This demo can transcribe up to {
|
|
|
|
|
34 |
)
|
35 |
|
36 |
if sr != SAMPLE_RATE:
|
37 |
data = librosa.resample(data, orig_sr=sr, target_sr=SAMPLE_RATE)
|
38 |
|
39 |
-
# monochannel
|
40 |
-
data = librosa.to_mono(data)
|
41 |
-
|
42 |
out_filename = os.path.join(tmpdir, utt_id + '.wav')
|
43 |
|
44 |
# save output audio
|
@@ -54,7 +76,6 @@ def transcribe(audio_filepath, src_lang, tgt_lang, pnc):
|
|
54 |
|
55 |
utt_id = uuid.uuid4()
|
56 |
with tempfile.TemporaryDirectory() as tmpdir:
|
57 |
-
|
58 |
converted_audio_filepath, duration = convert_audio(audio_filepath, tmpdir, str(utt_id))
|
59 |
|
60 |
# map src_lang and tgt_lang from long versions to short
|
@@ -102,9 +123,23 @@ def transcribe(audio_filepath, src_lang, tgt_lang, pnc):
|
|
102 |
fout.write(line + '\n')
|
103 |
|
104 |
# call transcribe, passing in manifest filepath
|
105 |
-
|
106 |
-
|
107 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
108 |
|
109 |
# add logic to make sure dropdown menus only suggest valid combos
|
110 |
def on_src_or_tgt_lang_change(src_lang_value, tgt_lang_value, pnc_value):
|
@@ -124,15 +159,15 @@ def on_src_or_tgt_lang_change(src_lang_value, tgt_lang_value, pnc_value):
|
|
124 |
tgt_lang_value, and then which states you can go to from there.
|
125 |
|
126 |
tgt lang
|
127 |
-
|
128 |
-
|
129 |
-
|
130 |
-
|
131 |
src ES| Y | Y | |
|
132 |
lang ------------------
|
133 |
-
|
134 |
-
|
135 |
-
|
136 |
"""
|
137 |
|
138 |
if src_lang_value == "English" and tgt_lang_value == "English":
|
|
|
6 |
import tempfile
|
7 |
import uuid
|
8 |
|
9 |
+
import torch
|
10 |
+
|
11 |
from nemo.collections.asr.models import ASRModel
|
12 |
+
from nemo.collections.asr.parts.utils.streaming_utils import FrameBatchMultiTaskAED
|
13 |
+
from nemo.collections.asr.parts.utils.transcribe_utils import get_buffered_pred_feat_multitaskAED
|
14 |
|
15 |
SAMPLE_RATE = 16000 # Hz
|
16 |
+
MAX_AUDIO_MINUTES = 10 # wont try to transcribe if longer than this
|
17 |
|
18 |
model = ASRModel.from_pretrained("nvidia/canary-1b")
|
19 |
model.eval()
|
20 |
|
21 |
+
# make sure beam size always 1 for consistency
|
22 |
+
model.change_decoding_strategy(None)
|
23 |
+
decoding_cfg = model.cfg.decoding
|
24 |
+
decoding_cfg.beam.beam_size = 1
|
25 |
+
model.change_decoding_strategy(decoding_cfg)
|
26 |
+
|
27 |
+
# setup for buffered inference
|
28 |
+
model.cfg.preprocessor.dither = 0.0
|
29 |
+
model.cfg.preprocessor.pad_to = 0
|
30 |
+
|
31 |
+
feature_stride = model.cfg.preprocessor['window_stride']
|
32 |
+
model_stride_in_secs = feature_stride * 8 # 8 = model stride, which is 8 for FastConformer
|
33 |
|
34 |
+
frame_asr = FrameBatchMultiTaskAED(
|
35 |
+
asr_model=model,
|
36 |
+
frame_len=40.0,
|
37 |
+
total_buffer=40.0,
|
38 |
+
batch_size=16,
|
39 |
+
)
|
40 |
|
41 |
+
amp_dtype = torch.float16
|
42 |
|
43 |
def convert_audio(audio_filepath, tmpdir, utt_id):
|
44 |
"""
|
|
|
47 |
Returns output filename and duration.
|
48 |
"""
|
49 |
|
50 |
+
data, sr = librosa.load(audio_filepath, sr=None, mono=False)
|
51 |
|
52 |
duration = librosa.get_duration(y=data, sr=sr)
|
53 |
|
54 |
+
if duration / 60.0 > MAX_AUDIO_MINUTES:
|
55 |
raise gr.Error(
|
56 |
+
f"This demo can transcribe up to {MAX_AUDIO_MINUTES} minutes of audio. "
|
57 |
+
"If you wish, you may trim the audio using the Audio viewer in Step 1 "
|
58 |
+
"(click on the scissors icon to start trimming audio)."
|
59 |
)
|
60 |
|
61 |
if sr != SAMPLE_RATE:
|
62 |
data = librosa.resample(data, orig_sr=sr, target_sr=SAMPLE_RATE)
|
63 |
|
|
|
|
|
|
|
64 |
out_filename = os.path.join(tmpdir, utt_id + '.wav')
|
65 |
|
66 |
# save output audio
|
|
|
76 |
|
77 |
utt_id = uuid.uuid4()
|
78 |
with tempfile.TemporaryDirectory() as tmpdir:
|
|
|
79 |
converted_audio_filepath, duration = convert_audio(audio_filepath, tmpdir, str(utt_id))
|
80 |
|
81 |
# map src_lang and tgt_lang from long versions to short
|
|
|
123 |
fout.write(line + '\n')
|
124 |
|
125 |
# call transcribe, passing in manifest filepath
|
126 |
+
if duration < 40:
|
127 |
+
output_text = model.transcribe(manifest_filepath)[0]
|
128 |
+
else: # do buffered inference
|
129 |
+
with torch.cuda.amp.autocast(dtype=amp_dtype): # TODO: make it work if no cuda
|
130 |
+
with torch.no_grad():
|
131 |
+
hyps = get_buffered_pred_feat_multitaskAED(
|
132 |
+
frame_asr,
|
133 |
+
model.cfg.preprocessor,
|
134 |
+
model_stride_in_secs,
|
135 |
+
model.device,
|
136 |
+
manifest=manifest_filepath,
|
137 |
+
filepaths=None,
|
138 |
+
)
|
139 |
+
|
140 |
+
output_text = hyps[0].text
|
141 |
+
|
142 |
+
return output_text
|
143 |
|
144 |
# add logic to make sure dropdown menus only suggest valid combos
|
145 |
def on_src_or_tgt_lang_change(src_lang_value, tgt_lang_value, pnc_value):
|
|
|
159 |
tgt_lang_value, and then which states you can go to from there.
|
160 |
|
161 |
tgt lang
|
162 |
+
- |EN |ES |FR |DE
|
163 |
+
------------------
|
164 |
+
EN| Y | Y | Y | Y
|
165 |
+
------------------
|
166 |
src ES| Y | Y | |
|
167 |
lang ------------------
|
168 |
+
FR| Y | | Y |
|
169 |
+
------------------
|
170 |
+
DE| Y | | | Y
|
171 |
"""
|
172 |
|
173 |
if src_lang_value == "English" and tgt_lang_value == "English":
|
requirements.txt
CHANGED
@@ -1 +1 @@
|
|
1 |
-
git+https://github.com/NVIDIA/NeMo.git@
|
|
|
1 |
+
git+https://github.com/NVIDIA/NeMo.git@61325fe0c70ef4294d8562991f6841d26b238e85#egg=nemo_toolkit[all] # commit from canary_buffer_infer branch
|