app.py

import gradio as gr
from transformers import AutoModelForCausalLM, AutoProcessor, TextIteratorStreamer
import librosa
from threading import Thread
import spaces

def split_audio(audio_arrays, chunk_limit=480000):
    CHUNK_LIM = chunk_limit
    audio_splits = []
    # Split the loaded audio to 30s chunks and extend the messages content
    for i in range(
        0,
        len(audio_arrays),
        CHUNK_LIM,
    ):
        audio_splits.append(audio_arrays[i : i + CHUNK_LIM])
    return audio_splits


def user(audio, text, chat_history):

    if audio is not None:
        chat_history.append(gr.ChatMessage(role="user", content={"path": audio, "alt_text": "Audio"}))
    chat_history.append({"role": "user", "content": text})
    return "", chat_history


@spaces.GPU
def process_audio(audio, text, chat_history):
    conversation = [
        {
            "role": "user",
            "content": [
            ],
        },
    ]
    audio_path = audio
    audio = librosa.load(audio, sr=16000)[0]

    if audio is not None:
        splitted_audio = split_audio(audio)
        for au in splitted_audio:
            conversation[0]["content"].append(
                {
                    "type": "audio_url",
                    "audio": "placeholder",
                }
            )
        # chat_history.append(gr.ChatMessage(role="user", content={"path": audio_path, "alt_text": "Audio"}))
    
    conversation[0]["content"].append(
        {
            "type": "text",
            "text": text,
        }
    )

    # chat_history.append({"role": "user", "content": text})
    # Set up the streamer for token generation
    streamer = TextIteratorStreamer(processor.tokenizer, skip_prompt=True, skip_special_tokens=True)
    prompt = processor.apply_chat_template(conversation, add_generation_prompt=True)
    inputs = processor(text=prompt, audios=splitted_audio, sampling_rate=16000, return_tensors="pt", padding=True)
    inputs = {k: v.to("cuda") for k, v in inputs.items()}
    # Set up generation arguments including max tokens and streamer
    generation_args = {
        "max_new_tokens": 4096,
        "streamer": streamer,
        "eos_token_id":151645, 
        "pad_token_id":151643,
        **inputs
    }
    # Start a separate thread for model generation to allow streaming output
    chat_history.append({"role": "assistant", "content": ""})
    thread = Thread(
        target=model.generate,
        kwargs=generation_args,
    )
    thread.start()
    for character in streamer:
        chat_history[-1]['content'] += character
        yield chat_history 

with gr.Blocks() as demo:
    gr.Markdown("## 🎙️ Aero-1-Audio")
    gr.Markdown(
    """
    ![logo](./examples/aero-1-audio.png)
    
    Aero-1-Audio is a lightweight audio-language model with only 1.5 billion parameters, trained on 50,000 hours of high-quality audio data. Despite its compact size, it supports a wide range of tasks, such as Automatic Speech Recognition (ASR), Basic Audio Understanding, Audio Instruction Following, and Scene Audio Analysis.

    Notably, Aero-1-Audio excels at lossless ASR on ultra-long audio—up to 16 minutes—without the need for audio segmentation.

    [Github](https://github.com/EvolvingLMMs-Lab/Aero-1/blob/main/README.md) | [Playground](https://huggingface.co/spaces/lmms-lab/Aero-1-Audio-Demo) | [Model Checkpoints](https://huggingface.co/lmms-lab/Aero-1-Audio-1.5B) | [Evaluation Results](https://github.com/EvolvingLMMs-Lab/lmms-eval/pull/658) | [Cookbook](https://www.lmms-lab.com/posts/lmms-lab-docs/aero_audio/)

    To explore its capabilities, you can upload your own audio or record your voice directly.
    Or simply start by trying the example demo below.

    ⚠️ Disclaimer: Aero-1-Audio is still under active development. Occasional inaccuracies may occur. We appreciate your understanding and welcome any feedback to help us make it better.
    """
    )

    chatbot = gr.Chatbot(type="messages")

    with gr.Row(variant="compact", equal_height=True):
        audio_input = gr.Audio(label="Speak Here", type="filepath")
        text_input = gr.Textbox(label="Text Input", placeholder="Please transcribe this audio for me", interactive=True)

    
    with gr.Row():
        chatbot_clear = gr.ClearButton([text_input, audio_input, chatbot], value="Clear")
        chatbot_submit = gr.Button("Submit", variant="primary")
        chatbot_submit.click(
            user,
            inputs=[audio_input, text_input, chatbot],
            outputs=[text_input, chatbot],
            queue=False
        ).then(
            process_audio,
            inputs=[audio_input, text_input, chatbot],
            outputs=[chatbot],
        )
    
    gr.Examples(
        [
            ["Please transcribe the audio for me", "./examples/elon_musk.mp3"],
            ["Please transcribe the audio for me", "./examples/nvidia_conference.mp3"],
            ["Please follow the instruction in the audio", "./examples/audio_instruction.wav"],
            ["What is the primary instrument featured in the solo of this track?", "./examples/music_under.wav"],
            ["What weather condition can be heard in the audio?", "./examples/audio_understand.wav"],
        ],
        inputs=[text_input, audio_input],
        label="Examples",
    )


if __name__ == "__main__":
    processor = AutoProcessor.from_pretrained("lmms-lab/Aero-1-Audio-1.5B", trust_remote_code=True)
    model = AutoModelForCausalLM.from_pretrained("lmms-lab/Aero-1-Audio-1.5B", device_map="cuda", torch_dtype="auto", attn_implementation="sdpa", trust_remote_code=True)
    demo.launch()