Shuka v1 is a language model which natively understands audio in Indic languages. It is an encoder-decoder model built by combining two models:
- Our state-of-the-art, in-house, audio encoder: Saaras v1
- Meta’s Llama3-8B-Instruct as the decoder
The encoder and decoder are connected by a small projector with ~60M parameters. During training, only the projector weights are finetuned while the rest of the network is frozen. Following our tradition of training models frugally, we train Shuka v1 on less than 100 hours of audio.
Though we only finetune the projector on English and Hindi data, the multilingual nature of our encoder makes Shuka v1 perform well on zero-shot QA in other Indic languages as well. We have tested on the model on Bengali, English, Gujarati, Hindi, Kannada, Malayalam, Marathi, Oriya, Punjabi, Tamil, and Telugu.
See what Shuka v1 can do in this demo video, and get started by using huggingface pipeline, as follows:
# install libraries
# pip install transformers==4.41.2 peft==0.11.1 librosa==0.10.2
import transformers
import librosa
# load the model pipeline on gpu:0
pipe = transformers.pipeline(model='sarvamai/shuka_v1', trust_remote_code=True, device=0, torch_dtype='bfloat16')
# get a sample audio
# wget https://huggingface.co/sarvamai/shuka_v1/resolve/main/hi-question.webm
audio, sr = librosa.load("./hi-question.webm", sr=16000)
turns = [
{'role': 'system', 'content': 'Respond naturally and informatively.'},
{'role': 'user', 'content': '<|audio|>'}
]
pipe({'audio': audio, 'turns': turns, 'sampling_rate': sr}, max_new_tokens=512)
For more details, please see our blog.