app.py

import os, logging, sys
import gradio as gr
import numpy as np
from groq import Groq
from dia.model import Dia



logging.basicConfig(
    level=logging.INFO,
    format="%(asctime)s [%(levelname)s] %(message)s",
    stream=sys.stdout
)

DEFAULT_REF_PATH = "https://storage.googleapis.com/chatterbox-demo-samples/prompts/female_shadowheart4.flac"
DEFAULT_GEN_TEXT = "Now let's make my mum's favourite. So three mars bars into the pan. Then we add the tuna and just stir for a bit, just let the chocolate and fish infuse. A sprinkle of olive oil and some tomato ketchup. Now smell that. Oh boy this is going to be incredible."
SAMPLES_PATH = os.path.join(os.getcwd(), "samples")
DEFAULT_REF_TEXT = "That place in the distance, it's huge and dedicated to Lady Shah. It can only mean one thing. I have a hidden place close to the cloister where night orchids bloom."

model = Dia.from_pretrained("nari-labs/Dia-1.6B-0626")

def transcribe(file_path: str):
    client = Groq()
    with open(file_path, "rb") as file:
        transcription = client.audio.transcriptions.create(
            file=(file_path, file.read()),
            model="whisper-large-v3-turbo",
            temperature=0,
            response_format="verbose_json",
        )

        if len(transcription.text) <= 0: logging.warn("Error while transcripting the reference audio.")
        else: logging.info(f"Transcribed: {transcription.text}")
        return transcription.text

def infer(
    gen_text: str,
    ref_text: str = DEFAULT_REF_TEXT,
    ref_audio_path: str = DEFAULT_REF_PATH,
) -> tuple[int, np.ndarray]:
    """
    Generates speech using NeuTTS-Air given a reference audio and text, and new text to synthesize.
    Args:
        gen_text (str): The new text to synthesize.
        ref_text (str): The text corresponding to the reference audio.
        ref_audio_path (str): The file path to the reference audio.
    Returns:
        tuple [int, np.ndarray]: A tuple containing the sample rate (44100) and the generated audio waveform as a numpy array.
    """

    if gen_text is None or not len(gen_text): 
        raise ValueError("Please insert the new text to synthesize.")
    if "female_shadowheart4.flac" not in ref_audio_path and ref_text == DEFAULT_REF_TEXT: ref_text = ""
    if not len(ref_text): 
        ref_text = transcribe(ref_audio_path)

    logging.info(f"Using reference: {ref_audio_path}")
    gr.Info("Starting inference request!")
    gr.Info("Encoding reference...")

    # ndarray[Unknown, Unknown] | list[ndarray[Unknown, Unknown]]
    output = model.generate(
        ref_text + gen_text,
        audio_prompt=ref_audio_path,
        use_torch_compile=False,
        verbose=True,
        cfg_scale=4.0,
        temperature=1.8,
        top_p=0.90,
        cfg_filter_top_k=50,
    )

    if isinstance(output, list):
        output = np.concatenate(output, axis=-1)  # Junta os pedaços de áudio
    elif not isinstance(output, np.ndarray):
        output = np.array(output, dtype=np.float32)

    return (44100, output)


demo = gr.Interface(
    fn=infer,
    inputs=[
        gr.Textbox(label="Text to Generate", value=DEFAULT_GEN_TEXT),
        gr.Textbox(label="Reference Text (Optional)", value=DEFAULT_REF_TEXT),
        gr.Audio(type="filepath", label="Reference Audio", value=DEFAULT_REF_PATH),
    ],
    outputs=gr.Audio(type="numpy", label="Generated Speech"),
    title="Dia-1.6B-0626",
    description="Upload a reference audio sample, provide the reference text, and enter new text to synthesize."
)

if __name__ == "__main__":
    demo.queue(max_size=10).launch(allowed_paths=[SAMPLES_PATH], mcp_server=False, inbrowser=True)