Update asr.py
Browse files
asr.py
CHANGED
@@ -1,9 +1,18 @@
|
|
1 |
import librosa
|
2 |
from transformers import Wav2Vec2ForCTC, AutoProcessor
|
3 |
import torch
|
|
|
|
|
|
|
|
|
4 |
|
5 |
ASR_SAMPLING_RATE = 16_000
|
6 |
|
|
|
|
|
|
|
|
|
|
|
7 |
|
8 |
MODEL_ID = "facebook/mms-1b-all"
|
9 |
|
@@ -11,21 +20,68 @@ processor = AutoProcessor.from_pretrained(MODEL_ID)
|
|
11 |
model = Wav2Vec2ForCTC.from_pretrained(MODEL_ID)
|
12 |
|
13 |
|
14 |
-
|
|
|
|
|
|
|
|
|
|
|
15 |
|
16 |
-
|
17 |
-
|
18 |
-
|
19 |
-
|
20 |
-
|
21 |
-
|
22 |
-
elif (microphone is None) and (file_upload is None):
|
23 |
-
return "ERROR: You have to either use the microphone or upload an audio file"
|
24 |
|
25 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
26 |
audio_samples = librosa.load(audio_fp, sr=ASR_SAMPLING_RATE, mono=True)[0]
|
27 |
|
28 |
-
lang_code = lang.split(
|
29 |
processor.tokenizer.set_target_lang(lang_code)
|
30 |
model.load_adapter(lang_code)
|
31 |
|
@@ -33,9 +89,42 @@ def transcribe(microphone, file_upload, lang):
|
|
33 |
audio_samples, sampling_rate=ASR_SAMPLING_RATE, return_tensors="pt"
|
34 |
)
|
35 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
36 |
with torch.no_grad():
|
37 |
outputs = model(**inputs).logits
|
38 |
|
39 |
-
|
40 |
-
|
41 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import librosa
|
2 |
from transformers import Wav2Vec2ForCTC, AutoProcessor
|
3 |
import torch
|
4 |
+
import json
|
5 |
+
|
6 |
+
from huggingface_hub import hf_hub_download
|
7 |
+
from torchaudio.models.decoder import ctc_decoder
|
8 |
|
9 |
ASR_SAMPLING_RATE = 16_000
|
10 |
|
11 |
+
ASR_LANGUAGES = {}
|
12 |
+
with open(f"data/asr/all_langs.tsv") as f:
|
13 |
+
for line in f:
|
14 |
+
iso, name = line.split(" ", 1)
|
15 |
+
ASR_LANGUAGES[iso] = name
|
16 |
|
17 |
MODEL_ID = "facebook/mms-1b-all"
|
18 |
|
|
|
20 |
model = Wav2Vec2ForCTC.from_pretrained(MODEL_ID)
|
21 |
|
22 |
|
23 |
+
# lm_decoding_config = {}
|
24 |
+
# lm_decoding_configfile = hf_hub_download(
|
25 |
+
# repo_id="facebook/mms-cclms",
|
26 |
+
# filename="decoding_config.json",
|
27 |
+
# subfolder="mms-1b-all",
|
28 |
+
# )
|
29 |
|
30 |
+
# with open(lm_decoding_configfile) as f:
|
31 |
+
# lm_decoding_config = json.loads(f.read())
|
32 |
+
|
33 |
+
# # allow language model decoding for "eng"
|
34 |
+
|
35 |
+
# decoding_config = lm_decoding_config["eng"]
|
|
|
|
|
36 |
|
37 |
+
# lm_file = hf_hub_download(
|
38 |
+
# repo_id="facebook/mms-cclms",
|
39 |
+
# filename=decoding_config["lmfile"].rsplit("/", 1)[1],
|
40 |
+
# subfolder=decoding_config["lmfile"].rsplit("/", 1)[0],
|
41 |
+
# )
|
42 |
+
# token_file = hf_hub_download(
|
43 |
+
# repo_id="facebook/mms-cclms",
|
44 |
+
# filename=decoding_config["tokensfile"].rsplit("/", 1)[1],
|
45 |
+
# subfolder=decoding_config["tokensfile"].rsplit("/", 1)[0],
|
46 |
+
# )
|
47 |
+
# lexicon_file = None
|
48 |
+
# if decoding_config["lexiconfile"] is not None:
|
49 |
+
# lexicon_file = hf_hub_download(
|
50 |
+
# repo_id="facebook/mms-cclms",
|
51 |
+
# filename=decoding_config["lexiconfile"].rsplit("/", 1)[1],
|
52 |
+
# subfolder=decoding_config["lexiconfile"].rsplit("/", 1)[0],
|
53 |
+
# )
|
54 |
+
|
55 |
+
# beam_search_decoder = ctc_decoder(
|
56 |
+
# lexicon=lexicon_file,
|
57 |
+
# tokens=token_file,
|
58 |
+
# lm=lm_file,
|
59 |
+
# nbest=1,
|
60 |
+
# beam_size=500,
|
61 |
+
# beam_size_token=50,
|
62 |
+
# lm_weight=float(decoding_config["lmweight"]),
|
63 |
+
# word_score=float(decoding_config["wordscore"]),
|
64 |
+
# sil_score=float(decoding_config["silweight"]),
|
65 |
+
# blank_token="<s>",
|
66 |
+
# )
|
67 |
+
|
68 |
+
|
69 |
+
def transcribe(
|
70 |
+
audio_source=None, microphone=None, file_upload=None, lang="eng (English)"
|
71 |
+
):
|
72 |
+
if type(microphone) is dict:
|
73 |
+
# HACK: microphone variable is a dict when running on examples
|
74 |
+
microphone = microphone["name"]
|
75 |
+
audio_fp = (
|
76 |
+
file_upload if "upload" in str(audio_source or "").lower() else microphone
|
77 |
+
)
|
78 |
+
|
79 |
+
if audio_fp is None:
|
80 |
+
return "ERROR: You have to either use the microphone or upload an audio file"
|
81 |
+
|
82 |
audio_samples = librosa.load(audio_fp, sr=ASR_SAMPLING_RATE, mono=True)[0]
|
83 |
|
84 |
+
lang_code = lang.split()[0]
|
85 |
processor.tokenizer.set_target_lang(lang_code)
|
86 |
model.load_adapter(lang_code)
|
87 |
|
|
|
89 |
audio_samples, sampling_rate=ASR_SAMPLING_RATE, return_tensors="pt"
|
90 |
)
|
91 |
|
92 |
+
# set device
|
93 |
+
if torch.cuda.is_available():
|
94 |
+
device = torch.device("cuda")
|
95 |
+
elif (
|
96 |
+
hasattr(torch.backends, "mps")
|
97 |
+
and torch.backends.mps.is_available()
|
98 |
+
and torch.backends.mps.is_built()
|
99 |
+
):
|
100 |
+
device = torch.device("mps")
|
101 |
+
else:
|
102 |
+
device = torch.device("cpu")
|
103 |
+
|
104 |
+
model.to(device)
|
105 |
+
inputs = inputs.to(device)
|
106 |
+
|
107 |
with torch.no_grad():
|
108 |
outputs = model(**inputs).logits
|
109 |
|
110 |
+
if lang_code != "eng" or True:
|
111 |
+
ids = torch.argmax(outputs, dim=-1)[0]
|
112 |
+
transcription = processor.decode(ids)
|
113 |
+
else:
|
114 |
+
assert False
|
115 |
+
# beam_search_result = beam_search_decoder(outputs.to("cpu"))
|
116 |
+
# transcription = " ".join(beam_search_result[0][0].words).strip()
|
117 |
+
|
118 |
+
return transcription
|
119 |
+
|
120 |
+
|
121 |
+
ASR_EXAMPLES = [
|
122 |
+
[None, "assets/english.mp3", None, "eng (English)"],
|
123 |
+
# [None, "assets/tamil.mp3", None, "tam (Tamil)"],
|
124 |
+
# [None, "assets/burmese.mp3", None, "mya (Burmese)"],
|
125 |
+
]
|
126 |
+
|
127 |
+
ASR_NOTE = """
|
128 |
+
The above demo doesn't use beam-search decoding using a language model.
|
129 |
+
Checkout the instructions [here](https://huggingface.co/facebook/mms-1b-all) on how to run LM decoding for better accuracy.
|
130 |
+
"""
|