Spaces:
Sleeping
Sleeping
| """ | |
| SonicMaster on HARP: text-controlled music restoration and mastering. | |
| The model runs here rather than on the authors' Space, so a request draws on | |
| this Space's own allowance instead of theirs. The weights come from | |
| amaai-lab/SonicMaster and the inference code is vendored from the authors' | |
| repository, which LICENSE-SonicMaster carries. | |
| """ | |
| import tempfile | |
| from pathlib import Path | |
| import gradio as gr | |
| import spaces | |
| from pyharp import (Channels, ModelCard, SampleRate, Subcategory, | |
| build_endpoint) | |
| import sonicmaster_runtime | |
| model_card = ModelCard( | |
| name="SonicMaster", | |
| description=( | |
| "Text-controlled all-in-one music restoration and mastering: one " | |
| "instruction in plain language fixes reverb, clipping, tonal balance, " | |
| "dynamics, or the stereo image. The model is a flow-matching " | |
| "transformer over Stable Audio's latent space, conditioned on a " | |
| "FLAN-T5 encoding of the instruction, so the wording of the prompt is " | |
| "what selects the treatment. Audio is resampled to 44.1 kHz stereo and " | |
| "processed in overlapping 30-second chunks, each conditioned on the " | |
| "tail of the one before it, which keeps a long track consistent across " | |
| "chunk boundaries. SonicMaster itself is Apache 2.0, but it decodes " | |
| "through Stable Audio Open's VAE, so results are also governed by the " | |
| "Stability AI Community License: " | |
| "https://huggingface.co/stabilityai/stable-audio-open-1.0/blob/main/LICENSE.md " | |
| "which permits non-commercial use and commercial use below an annual " | |
| "revenue threshold. Repository: " | |
| "https://github.com/AMAAI-Lab/SonicMaster Paper: " | |
| "https://arxiv.org/abs/2508.03448" | |
| ), | |
| author="Jan Melechovsky, Ambuj Mehrish, Dorien Herremans", | |
| tags=[Subcategory.POST_PROCESSING, SampleRate(sonicmaster_runtime.SAMPLE_RATE), | |
| Channels(sonicmaster_runtime.CHANNELS), "text-guided"], | |
| ) | |
| def process_fn(input_audio_path, prompt, num_inference_steps, guidance_scale, | |
| solver, seed): | |
| """ | |
| Apply the instruction to the recording. | |
| Args: | |
| input_audio_path (str): The recording to edit. | |
| prompt (str): What to change about it. | |
| num_inference_steps (float): Flow solver steps per chunk. | |
| guidance_scale (float): How strongly the instruction is followed. | |
| solver (str): Which solver steps the flow. | |
| seed (float): Seed for the noise. | |
| Returns: | |
| output_path (str): The edited audio. | |
| """ | |
| if not (prompt or "").strip(): | |
| raise gr.Error("Describe what to change about the audio.") | |
| output_path = Path(tempfile.mkdtemp()) / "sonicmaster.wav" | |
| return sonicmaster_runtime.enhance( | |
| input_audio_path, | |
| prompt.strip(), | |
| output_path, | |
| num_inference_steps=int(num_inference_steps), | |
| guidance_scale=float(guidance_scale), | |
| seed=int(seed), | |
| solver=solver, | |
| ) | |
| if __name__ == "__main__": | |
| with gr.Blocks() as demo: | |
| input_components = [ | |
| gr.Audio(type="filepath", label="Input audio").harp_required(True).set_info(( | |
| "The recording to treat. It is resampled to 44.1 kHz and " | |
| "forced to stereo, since that is what the model was trained " | |
| "on, and a mono input is duplicated across both channels." | |
| )), | |
| gr.Textbox(label="Instruction", value="Remove the reverb from this recording.", | |
| info=( | |
| "What to change, in plain language, for example " | |
| "\"reduce the reverb\", \"brighten the high end\", " | |
| "\"even out the loudness\", or \"widen the stereo " | |
| "image\". One instruction at a time works better " | |
| "than several combined." | |
| )), | |
| gr.Slider(minimum=1, maximum=50, step=1, value=10, | |
| label="Solver steps", | |
| info=( | |
| "How many steps the flow solver takes per chunk. " | |
| "More steps follow the instruction more closely and " | |
| "cost proportionally more time." | |
| )), | |
| gr.Slider(minimum=1.0, maximum=7.0, step=0.1, value=1.0, | |
| label="Instruction strength", | |
| info=( | |
| "How strongly the instruction is weighted against " | |
| "leaving the audio alone. Raising it applies a more " | |
| "pronounced treatment, at the risk of artifacts." | |
| )), | |
| gr.Dropdown(choices=["Euler", "rk4"], value="Euler", label="Solver", | |
| info=( | |
| "Euler takes one evaluation per step. rk4 takes " | |
| "four, which is steadier at a low step count and " | |
| "roughly four times slower." | |
| )), | |
| gr.Number(value=0, precision=0, label="Seed", | |
| info=( | |
| "Seeds the noise the solver starts from, so the same " | |
| "seed and settings reproduce a result." | |
| )), | |
| ] | |
| output_components = [ | |
| gr.Audio(type="filepath", label="Treated audio").set_info(( | |
| "The result, at 44.1 kHz stereo." | |
| )), | |
| ] | |
| build_endpoint( | |
| model_card=model_card, | |
| input_components=input_components, | |
| output_components=output_components, | |
| process_fn=process_fn, | |
| ) | |
| demo.queue().launch(server_name="0.0.0.0", server_port=7860, | |
| show_error=True, pwa=True) | |