Spaces:
Running
Running
File size: 2,174 Bytes
07c5b0a 3f8984b 07c5b0a e758b0c 07c5b0a 3f8984b 8d1a8e1 24e6612 3f8984b 24e6612 3f8984b 8d1a8e1 24e6612 8d1a8e1 3f8984b 3dd81aa |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 |
import librosa
import torch
import numpy as np
import langid # Language detection library
from transformers import Wav2Vec2ForCTC, AutoProcessor
ASR_SAMPLING_RATE = 16_000
MODEL_ID = "facebook/mms-1b-all"
# openai/whisper-large-v3-turbo
#ASR_SAMPLING_RATE = 16_000
#MODEL_ID = "openai/whisper-large-v3-turbo"
# Load MMS Model
processor = AutoProcessor.from_pretrained(MODEL_ID)
model = Wav2Vec2ForCTC.from_pretrained(MODEL_ID)
model.eval()
def detect_language(text):
"""Detects language using langid (fast & lightweight)."""
lang, _ = langid.classify(text)
return lang if lang in ["en", "sw"] else "en" # Default to English
def transcribe_auto(audio_data=None):
if not audio_data:
return "<<ERROR: Empty Audio Input>>"
# Process Microphone Input
if isinstance(audio_data, tuple):
sr, audio_samples = audio_data
audio_samples = (audio_samples / 32768.0).astype(np.float32)
if sr != ASR_SAMPLING_RATE:
audio_samples = librosa.resample(audio_samples, orig_sr=sr, target_sr=ASR_SAMPLING_RATE)
# Process File Upload Input
else:
if not isinstance(audio_data, str):
return "<<ERROR: Invalid Audio Input>>"
audio_samples = librosa.load(audio_data, sr=ASR_SAMPLING_RATE, mono=True)[0]
inputs = processor(audio_samples, sampling_rate=ASR_SAMPLING_RATE, return_tensors="pt")
# **Step 1: Transcribe without Language Detection**
with torch.no_grad():
outputs = model(**inputs).logits
ids = torch.argmax(outputs, dim=-1)[0]
raw_transcription = processor.decode(ids)
# **Step 2: Detect Language from Transcription**
detected_lang = detect_language(raw_transcription)
lang_code = "eng" if detected_lang == "en" else "swh"
# **Step 3: Reload Model with Correct Adapter**
processor.tokenizer.set_target_lang(lang_code)
model.load_adapter(lang_code)
# **Step 4: Transcribe Again with Correct Adapter**
with torch.no_grad():
outputs = model(**inputs).logits
ids = torch.argmax(outputs, dim=-1)[0]
final_transcription = processor.decode(ids)
return f"{final_transcription}" |