Spaces:
Running
Running
Refactor ASR module; add Paraformer support
Browse files- config/interface/options.yaml +2 -4
- modules/asr/__init__.py +11 -0
- modules/asr/base.py +24 -0
- modules/asr/paraformer.py +86 -0
- modules/asr/registry.py +19 -0
- modules/asr/whisper.py +82 -0
- tests/test_asr_infer.py +19 -0
config/interface/options.yaml
CHANGED
@@ -7,10 +7,8 @@ asr_models:
|
|
7 |
name: Whisper medium
|
8 |
- id: openai/whisper-small
|
9 |
name: Whisper small
|
10 |
-
- id:
|
11 |
-
name:
|
12 |
-
- id: facebook/wav2vec2-base-960h
|
13 |
-
name: Wav2Vec2-Base-960h
|
14 |
|
15 |
llm_models:
|
16 |
- id: gemini-2.5-flash
|
|
|
7 |
name: Whisper medium
|
8 |
- id: openai/whisper-small
|
9 |
name: Whisper small
|
10 |
+
- id: funasr/paraformer-zh
|
11 |
+
name: Paraformer-zh
|
|
|
|
|
12 |
|
13 |
llm_models:
|
14 |
- id: gemini-2.5-flash
|
modules/asr/__init__.py
ADDED
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from .base import AbstractASRModel
|
2 |
+
from .registry import ASR_MODEL_REGISTRY, get_asr_model, register_asr_model
|
3 |
+
from .whisper import WhisperASR
|
4 |
+
from .paraformer import ParaformerASR
|
5 |
+
|
6 |
+
__all__ = [
|
7 |
+
"AbstractASRModel",
|
8 |
+
"get_asr_model",
|
9 |
+
"register_asr_model",
|
10 |
+
"ASR_MODEL_REGISTRY",
|
11 |
+
]
|
modules/asr/base.py
ADDED
@@ -0,0 +1,24 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from abc import ABC, abstractmethod
|
2 |
+
from typing import Optional
|
3 |
+
|
4 |
+
import numpy as np
|
5 |
+
|
6 |
+
|
7 |
+
class AbstractASRModel(ABC):
|
8 |
+
def __init__(
|
9 |
+
self, model_id: str, device: str = "cpu", cache_dir: str = "cache", **kwargs
|
10 |
+
):
|
11 |
+
print(f"Loading ASR model {model_id}...")
|
12 |
+
self.model_id = model_id
|
13 |
+
self.device = device
|
14 |
+
self.cache_dir = cache_dir
|
15 |
+
|
16 |
+
@abstractmethod
|
17 |
+
def transcribe(
|
18 |
+
self,
|
19 |
+
audio: np.ndarray,
|
20 |
+
audio_sample_rate: int,
|
21 |
+
language: Optional[str] = None,
|
22 |
+
**kwargs,
|
23 |
+
) -> str:
|
24 |
+
pass
|
modules/asr/paraformer.py
ADDED
@@ -0,0 +1,86 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import tempfile
|
3 |
+
from typing import Optional
|
4 |
+
|
5 |
+
import numpy as np
|
6 |
+
import soundfile as sf
|
7 |
+
|
8 |
+
try:
|
9 |
+
from funasr import AutoModel
|
10 |
+
except ImportError:
|
11 |
+
AutoModel = None
|
12 |
+
|
13 |
+
from .base import AbstractASRModel
|
14 |
+
from .registry import register_asr_model
|
15 |
+
|
16 |
+
|
17 |
+
@register_asr_model("funasr/paraformer-zh")
|
18 |
+
class ParaformerASR(AbstractASRModel):
|
19 |
+
def __init__(
|
20 |
+
self, model_id: str, device: str = "cpu", cache_dir: str = "cache", **kwargs
|
21 |
+
):
|
22 |
+
super().__init__(model_id, device, cache_dir, **kwargs)
|
23 |
+
|
24 |
+
if AutoModel is None:
|
25 |
+
raise ImportError(
|
26 |
+
"funasr is not installed. Please install it with: pip3 install -U funasr"
|
27 |
+
)
|
28 |
+
|
29 |
+
model_name = model_id.replace("funasr/", "")
|
30 |
+
language = model_name.split("-")[1]
|
31 |
+
if language == "zh":
|
32 |
+
self.language = "mandarin"
|
33 |
+
elif language == "en":
|
34 |
+
self.language = "english"
|
35 |
+
else:
|
36 |
+
raise ValueError(
|
37 |
+
f"Language cannot be determined. {model_id} is not supported"
|
38 |
+
)
|
39 |
+
|
40 |
+
try:
|
41 |
+
original_cache_dir = os.getenv("MODELSCOPE_CACHE")
|
42 |
+
os.makedirs(cache_dir, exist_ok=True)
|
43 |
+
os.environ["MODELSCOPE_CACHE"] = cache_dir
|
44 |
+
self.model = AutoModel(
|
45 |
+
model=model_name,
|
46 |
+
model_revision="v2.0.4",
|
47 |
+
vad_model="fsmn-vad",
|
48 |
+
vad_model_revision="v2.0.4",
|
49 |
+
punc_model="ct-punc-c",
|
50 |
+
punc_model_revision="v2.0.4",
|
51 |
+
device=device,
|
52 |
+
)
|
53 |
+
if original_cache_dir:
|
54 |
+
os.environ["MODELSCOPE_CACHE"] = original_cache_dir
|
55 |
+
else:
|
56 |
+
del os.environ["MODELSCOPE_CACHE"]
|
57 |
+
|
58 |
+
except Exception as e:
|
59 |
+
raise ValueError(f"Error loading Paraformer model: {e}")
|
60 |
+
|
61 |
+
def transcribe(
|
62 |
+
self,
|
63 |
+
audio: np.ndarray,
|
64 |
+
audio_sample_rate: int,
|
65 |
+
language: Optional[str] = None,
|
66 |
+
**kwargs,
|
67 |
+
) -> str:
|
68 |
+
if language and language != self.language:
|
69 |
+
raise ValueError(
|
70 |
+
f"Paraformer model {self.model_id} only supports {self.language} language, but {language} was requested"
|
71 |
+
)
|
72 |
+
|
73 |
+
try:
|
74 |
+
with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as f:
|
75 |
+
sf.write(f.name, audio, audio_sample_rate)
|
76 |
+
temp_file = f.name
|
77 |
+
|
78 |
+
result = self.model.generate(input=temp_file, batch_size_s=300, **kwargs)
|
79 |
+
|
80 |
+
os.unlink(temp_file)
|
81 |
+
|
82 |
+
print(f"Transcription result: {result}, type: {type(result)}")
|
83 |
+
|
84 |
+
return result[0]["text"]
|
85 |
+
except Exception as e:
|
86 |
+
raise ValueError(f"Error during transcription: {e}")
|
modules/asr/registry.py
ADDED
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from .base import AbstractASRModel
|
2 |
+
|
3 |
+
ASR_MODEL_REGISTRY = {}
|
4 |
+
|
5 |
+
|
6 |
+
def register_asr_model(prefix: str):
|
7 |
+
def wrapper(cls):
|
8 |
+
assert issubclass(cls, AbstractASRModel), f"{cls} must inherit AbstractASRModel"
|
9 |
+
ASR_MODEL_REGISTRY[prefix] = cls
|
10 |
+
return cls
|
11 |
+
|
12 |
+
return wrapper
|
13 |
+
|
14 |
+
|
15 |
+
def get_asr_model(model_id: str, device="cpu", **kwargs) -> AbstractASRModel:
|
16 |
+
for prefix, cls in ASR_MODEL_REGISTRY.items():
|
17 |
+
if model_id.startswith(prefix):
|
18 |
+
return cls(model_id, device=device, **kwargs)
|
19 |
+
raise ValueError(f"No ASR wrapper found for model: {model_id}")
|
modules/asr/whisper.py
ADDED
@@ -0,0 +1,82 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
from typing import Optional
|
3 |
+
|
4 |
+
import librosa
|
5 |
+
import numpy as np
|
6 |
+
from transformers.pipelines import pipeline
|
7 |
+
|
8 |
+
from .base import AbstractASRModel
|
9 |
+
from .registry import register_asr_model
|
10 |
+
|
11 |
+
hf_token = os.getenv("HF_TOKEN")
|
12 |
+
|
13 |
+
|
14 |
+
@register_asr_model("openai/whisper")
|
15 |
+
class WhisperASR(AbstractASRModel):
|
16 |
+
def __init__(
|
17 |
+
self, model_id: str, device: str = "cpu", cache_dir: str = "cache", **kwargs
|
18 |
+
):
|
19 |
+
super().__init__(model_id, device, cache_dir, **kwargs)
|
20 |
+
model_kwargs = kwargs.setdefault("model_kwargs", {})
|
21 |
+
model_kwargs["cache_dir"] = cache_dir
|
22 |
+
self.pipe = pipeline(
|
23 |
+
"automatic-speech-recognition",
|
24 |
+
model=model_id,
|
25 |
+
device=0 if device == "cuda" else -1,
|
26 |
+
token=hf_token,
|
27 |
+
**kwargs,
|
28 |
+
)
|
29 |
+
|
30 |
+
def transcribe(
|
31 |
+
self,
|
32 |
+
audio: np.ndarray,
|
33 |
+
audio_sample_rate: int,
|
34 |
+
language: Optional[str] = None,
|
35 |
+
**kwargs,
|
36 |
+
) -> str:
|
37 |
+
"""
|
38 |
+
Transcribe audio using Whisper model
|
39 |
+
|
40 |
+
Args:
|
41 |
+
audio: Audio numpy array
|
42 |
+
audio_sample_rate: Sample rate of the audio
|
43 |
+
language: Language hint (optional)
|
44 |
+
|
45 |
+
Returns:
|
46 |
+
Transcribed text as string
|
47 |
+
"""
|
48 |
+
try:
|
49 |
+
# Resample to 16kHz if needed
|
50 |
+
if audio_sample_rate != 16000:
|
51 |
+
audio = librosa.resample(
|
52 |
+
audio, orig_sr=audio_sample_rate, target_sr=16000
|
53 |
+
)
|
54 |
+
|
55 |
+
# Generate transcription
|
56 |
+
generate_kwargs = {}
|
57 |
+
if language:
|
58 |
+
generate_kwargs["language"] = language
|
59 |
+
|
60 |
+
result = self.pipe(
|
61 |
+
audio,
|
62 |
+
generate_kwargs=generate_kwargs,
|
63 |
+
return_timestamps=False,
|
64 |
+
**kwargs,
|
65 |
+
)
|
66 |
+
|
67 |
+
# Extract text from result
|
68 |
+
if isinstance(result, dict) and "text" in result:
|
69 |
+
return result["text"]
|
70 |
+
elif isinstance(result, list) and len(result) > 0:
|
71 |
+
# Handle list of results
|
72 |
+
first_result = result[0]
|
73 |
+
if isinstance(first_result, dict):
|
74 |
+
return first_result.get("text", str(first_result))
|
75 |
+
else:
|
76 |
+
return str(first_result)
|
77 |
+
else:
|
78 |
+
return str(result)
|
79 |
+
|
80 |
+
except Exception as e:
|
81 |
+
print(f"Error during Whisper transcription: {e}")
|
82 |
+
return ""
|
tests/test_asr_infer.py
ADDED
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from modules.asr import get_asr_model
|
2 |
+
import librosa
|
3 |
+
|
4 |
+
if __name__ == "__main__":
|
5 |
+
supported_asrs = [
|
6 |
+
"funasr/paraformer-zh",
|
7 |
+
"openai/whisper-large-v3-turbo",
|
8 |
+
]
|
9 |
+
for model_id in supported_asrs:
|
10 |
+
try:
|
11 |
+
print(f"Loading model: {model_id}")
|
12 |
+
asr_model = get_asr_model(model_id, device="cpu", cache_dir=".cache")
|
13 |
+
audio, sample_rate = librosa.load("tests/audio/hello.wav", sr=None)
|
14 |
+
result = asr_model.transcribe(audio, sample_rate, language="mandarin")
|
15 |
+
print(result)
|
16 |
+
except Exception as e:
|
17 |
+
print(f"Failed to load model {model_id}: {e}")
|
18 |
+
breakpoint()
|
19 |
+
continue
|