import os import gradio as gr import numpy as np import torch from groq import Groq from transformers import pipeline from transformers.utils import is_flash_attn_2_available transcriber = pipeline("automatic-speech-recognition", model="openai/whisper-large-v3", torch_dtype=torch.float16, device="cuda:0", model_kwargs={"attn_implementation": "flash_attention_2"} if is_flash_attn_2_available() else {"attn_implementation": "sdpa"}, ) groq_client = Groq(api_key=os.getenv('GROQ_API_KEY')) def transcribe(stream, new_chunk): """ Transcribes using whisper """ sr, y = new_chunk # Convert stereo to mono if necessary if y.ndim == 2 and y.shape[1] == 2: y = y.mean(axis=1) # Averaging both channels if stereo y = y.astype(np.float32) # Normalization y /= np.max(np.abs(y)) if stream is not None: stream = np.concatenate([stream, y]) else: stream = y return stream, transcriber({"sampling_rate": sr, "raw": stream})["text"] def autocomplete(text): """ Autocomplete the text using Gemma. """ if text != "": response = groq_client.chat.completions.create( model='gemma-7b-it', messages=[{"role": "system", "content": "You are a friendly assistant named Gemma."}, {"role": "user", "content": text}] ) return response.choices[0].message.content def process_audio(input_audio, new_chunk): """ Process the audio input by transcribing and completing the sentences. Accumulate results to return to Gradio interface. """ stream, transcription = transcribe(input_audio, new_chunk) text = autocomplete(transcription) print (transcription, text) return stream, text demo = gr.Interface( fn = process_audio, inputs = ["state", gr.Audio(sources=["microphone"], streaming=True)], outputs = ["state", gr.Markdown()], title="Hey Gemma ☎️", description="Powered by [whisper-base-en](https://huggingface.co/openai/whisper-base.en), and [gemma-7b-it](https://huggingface.co/google/gemma-7b-it) (via [Groq](https://groq.com/))", live=True, allow_flagging="never" ) demo.launch()