Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -39,8 +39,10 @@ This Gradio demo showcases **IndicSeamlessM4T**, a fine-tuned **SeamlessM4T** mo
|
|
39 |
"""
|
40 |
|
41 |
hf_token = os.getenv("HF_TOKEN")
|
|
|
|
|
42 |
|
43 |
-
model = SeamlessM4Tv2ForSpeechToText.from_pretrained("ai4bharat/seamless-m4t-v2-large-stt", torch_dtype=
|
44 |
processor = SeamlessM4TFeatureExtractor.from_pretrained("ai4bharat/seamless-m4t-v2-large-stt", token=hf_token)
|
45 |
tokenizer = SeamlessM4TTokenizer.from_pretrained("ai4bharat/seamless-m4t-v2-large-stt", token=hf_token)
|
46 |
|
@@ -50,17 +52,6 @@ AUDIO_SAMPLE_RATE = 16000.0
|
|
50 |
MAX_INPUT_AUDIO_LENGTH = 60 # in seconds
|
51 |
DEFAULT_TARGET_LANGUAGE = "Hindi"
|
52 |
|
53 |
-
if torch.cuda.is_available():
|
54 |
-
device = torch.device("cuda:0")
|
55 |
-
dtype = torch.float16
|
56 |
-
else:
|
57 |
-
device = torch.device("cpu")
|
58 |
-
dtype = torch.float32
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
def preprocess_audio(input_audio: str) -> None:
|
65 |
arr, org_sr = torchaudio.load(input_audio)
|
66 |
new_arr = torchaudio.functional.resample(arr, orig_freq=org_sr, new_freq=AUDIO_SAMPLE_RATE)
|
@@ -78,7 +69,7 @@ def run_s2tt(input_audio: str, source_language: str, target_language: str) -> st
|
|
78 |
|
79 |
input_audio, orig_freq = torchaudio.load(input_audio)
|
80 |
input_audio = torchaudio.functional.resample(input_audio, orig_freq=orig_freq, new_freq=16000)
|
81 |
-
audio_inputs= processor(input_audio, sampling_rate=16000, return_tensors="pt").to(device="cuda",dtype=
|
82 |
|
83 |
text_out = model.generate(**audio_inputs, tgt_lang=target_language_code)[0].float().cpu().numpy().squeeze()
|
84 |
|
@@ -91,7 +82,7 @@ def run_asr(input_audio: str, target_language: str) -> str:
|
|
91 |
|
92 |
input_audio, orig_freq = torchaudio.load(input_audio)
|
93 |
input_audio = torchaudio.functional.resample(input_audio, orig_freq=orig_freq, new_freq=16000)
|
94 |
-
audio_inputs= processor(input_audio, sampling_rate=16000, return_tensors="pt").to(device="cuda",dtype=
|
95 |
|
96 |
text_out = model.generate(**audio_inputs, tgt_lang=target_language_code)[0].float().cpu().numpy().squeeze()
|
97 |
|
|
|
39 |
"""
|
40 |
|
41 |
hf_token = os.getenv("HF_TOKEN")
|
42 |
+
device = "cuda:0" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"
|
43 |
+
torch_dtype = torch.bfloat16 if device != "cpu" else torch.float32
|
44 |
|
45 |
+
model = SeamlessM4Tv2ForSpeechToText.from_pretrained("ai4bharat/seamless-m4t-v2-large-stt", torch_dtype=torch_dtype, token=hf_token).to(device)
|
46 |
processor = SeamlessM4TFeatureExtractor.from_pretrained("ai4bharat/seamless-m4t-v2-large-stt", token=hf_token)
|
47 |
tokenizer = SeamlessM4TTokenizer.from_pretrained("ai4bharat/seamless-m4t-v2-large-stt", token=hf_token)
|
48 |
|
|
|
52 |
MAX_INPUT_AUDIO_LENGTH = 60 # in seconds
|
53 |
DEFAULT_TARGET_LANGUAGE = "Hindi"
|
54 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
55 |
def preprocess_audio(input_audio: str) -> None:
|
56 |
arr, org_sr = torchaudio.load(input_audio)
|
57 |
new_arr = torchaudio.functional.resample(arr, orig_freq=org_sr, new_freq=AUDIO_SAMPLE_RATE)
|
|
|
69 |
|
70 |
input_audio, orig_freq = torchaudio.load(input_audio)
|
71 |
input_audio = torchaudio.functional.resample(input_audio, orig_freq=orig_freq, new_freq=16000)
|
72 |
+
audio_inputs= processor(input_audio, sampling_rate=16000, return_tensors="pt").to(device="cuda", dtype=torch_dtype)
|
73 |
|
74 |
text_out = model.generate(**audio_inputs, tgt_lang=target_language_code)[0].float().cpu().numpy().squeeze()
|
75 |
|
|
|
82 |
|
83 |
input_audio, orig_freq = torchaudio.load(input_audio)
|
84 |
input_audio = torchaudio.functional.resample(input_audio, orig_freq=orig_freq, new_freq=16000)
|
85 |
+
audio_inputs= processor(input_audio, sampling_rate=16000, return_tensors="pt").to(device="cuda", dtype=torch_dtype)
|
86 |
|
87 |
text_out = model.generate(**audio_inputs, tgt_lang=target_language_code)[0].float().cpu().numpy().squeeze()
|
88 |
|