Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -28,7 +28,7 @@ device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
|
|
| 28 |
print(f"Bruker enhet: {device}")
|
| 29 |
|
| 30 |
@spaces.GPU(duration=60 * 2)
|
| 31 |
-
def pipe(file, return_timestamps=False,
|
| 32 |
asr = pipeline(
|
| 33 |
task="automatic-speech-recognition",
|
| 34 |
model=MODEL_NAME,
|
|
@@ -38,18 +38,11 @@ def pipe(file, return_timestamps=False,lang_nn=False):
|
|
| 38 |
torch_dtype=torch.float16,
|
| 39 |
model_kwargs={"attn_implementation": "flash_attention_2", "num_beams": 5} if FLASH_ATTENTION else {"attn_implementation": "sdpa", "num_beams": 5},
|
| 40 |
)
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
)
|
| 47 |
-
else:
|
| 48 |
-
asr.model.config.forced_decoder_ids = asr.tokenizer.get_decoder_prompt_ids(
|
| 49 |
-
language="nn",
|
| 50 |
-
task="transcribe",
|
| 51 |
-
no_timestamps=not return_timestamps,
|
| 52 |
-
)
|
| 53 |
return asr(file, return_timestamps=return_timestamps, batch_size=24)
|
| 54 |
|
| 55 |
def format_output(text):
|
|
@@ -81,10 +74,10 @@ def transcribe(file, return_timestamps=False,lang_nn=False):
|
|
| 81 |
|
| 82 |
if not lang_nn:
|
| 83 |
if not return_timestamps:
|
| 84 |
-
text = pipe(file_to_transcribe)["text"]
|
| 85 |
formatted_text = format_output(text)
|
| 86 |
else:
|
| 87 |
-
chunks = pipe(file_to_transcribe, return_timestamps=True)["chunks"]
|
| 88 |
text = []
|
| 89 |
for chunk in chunks:
|
| 90 |
start_time = time.strftime('%H:%M:%S', time.gmtime(chunk["timestamp"][0])) if chunk["timestamp"][0] is not None else "??:??:??"
|
|
@@ -94,10 +87,10 @@ def transcribe(file, return_timestamps=False,lang_nn=False):
|
|
| 94 |
formatted_text = "<br>".join(text)
|
| 95 |
else:
|
| 96 |
if not return_timestamps:
|
| 97 |
-
text = pipe(file_to_transcribe,
|
| 98 |
formatted_text = format_output(text)
|
| 99 |
else:
|
| 100 |
-
chunks = pipe(file_to_transcribe, return_timestamps=True,
|
| 101 |
text = []
|
| 102 |
for chunk in chunks:
|
| 103 |
start_time = time.strftime('%H:%M:%S', time.gmtime(chunk["timestamp"][0])) if chunk["timestamp"][0] is not None else "??:??:??"
|
|
|
|
| 28 |
print(f"Bruker enhet: {device}")
|
| 29 |
|
| 30 |
@spaces.GPU(duration=60 * 2)
|
| 31 |
+
def pipe(file, return_timestamps=False,lang):
|
| 32 |
asr = pipeline(
|
| 33 |
task="automatic-speech-recognition",
|
| 34 |
model=MODEL_NAME,
|
|
|
|
| 38 |
torch_dtype=torch.float16,
|
| 39 |
model_kwargs={"attn_implementation": "flash_attention_2", "num_beams": 5} if FLASH_ATTENTION else {"attn_implementation": "sdpa", "num_beams": 5},
|
| 40 |
)
|
| 41 |
+
asr.model.config.forced_decoder_ids = asr.tokenizer.get_decoder_prompt_ids(
|
| 42 |
+
language=lang,
|
| 43 |
+
task="transcribe",
|
| 44 |
+
no_timestamps=not return_timestamps,
|
| 45 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 46 |
return asr(file, return_timestamps=return_timestamps, batch_size=24)
|
| 47 |
|
| 48 |
def format_output(text):
|
|
|
|
| 74 |
|
| 75 |
if not lang_nn:
|
| 76 |
if not return_timestamps:
|
| 77 |
+
text = pipe(file_to_transcribe,lang="no")["text"]
|
| 78 |
formatted_text = format_output(text)
|
| 79 |
else:
|
| 80 |
+
chunks = pipe(file_to_transcribe, return_timestamps=True,lang="no")["chunks"]
|
| 81 |
text = []
|
| 82 |
for chunk in chunks:
|
| 83 |
start_time = time.strftime('%H:%M:%S', time.gmtime(chunk["timestamp"][0])) if chunk["timestamp"][0] is not None else "??:??:??"
|
|
|
|
| 87 |
formatted_text = "<br>".join(text)
|
| 88 |
else:
|
| 89 |
if not return_timestamps:
|
| 90 |
+
text = pipe(file_to_transcribe,lang="nn")["text"]
|
| 91 |
formatted_text = format_output(text)
|
| 92 |
else:
|
| 93 |
+
chunks = pipe(file_to_transcribe, return_timestamps=True,lang="nn")["chunks"]
|
| 94 |
text = []
|
| 95 |
for chunk in chunks:
|
| 96 |
start_time = time.strftime('%H:%M:%S', time.gmtime(chunk["timestamp"][0])) if chunk["timestamp"][0] is not None else "??:??:??"
|