SoundStorm: Parallel Audio Generation

[vc_row][vc_column][vc_headings linewidth=”0″ borderwidth=”1″ borderclr=”#000000″ title=”SoundStorm” google_fonts=”font_family:Comfortaa%3A300%2Cregular%2C700|font_style:700%20bold%20regular%3A700%3Anormal” titlesize=”60″ titleclr=”#000000″]Parallel Audio Generation[/vc_headings][vc_single_image image=”3056″ alignment=”center” onclick=”custom_link” img_link_target=”_blank”][vc_column_text]SoundStorm is a model for efficient, non-autoregressive audio generation. It takes the semantic tokens of AudioLM as input and uses bidirectional attention and confidence-based parallel decoding to generate the tokens of a neural audio codec. SoundStorm can produce high-quality audio faster and more consistently than the autoregressive approach of AudioLM. It can also synthesize natural dialogue segments from a transcript with speaker turns and voice prompts.[/vc_column_text][vc_btn title=”Github Page” color=”inverse” align=”center” i_align=”right” i_icon_fontawesome=”fab fa-github” add_icon=”true” link=”url:https%3A%2F%2Fgithub.com%2Flucidrains%2Fsoundstorm-pytorch|title:Github%20Page”][vc_btn title=”Project Page” color=”inverse” align=”center” i_align=”right” i_icon_fontawesome=”fab fa-github-square” add_icon=”true” link=”url:https%3A%2F%2Fgoogle-research.github.io%2Fseanet%2Fsoundstorm%2Fexamples%2F|title:Github%20Page”][vc_separator][/vc_column][/vc_row][vc_row][vc_column][vc_headings style=”theme4″ borderclr=”#000000″ style2=”image” title=”Soundstorm – Pytorch (wip)” google_fonts=”font_family:Comfortaa%3A300%2Cregular%2C700|font_style:700%20bold%20regular%3A700%3Anormal” lineheight=”3″ titlesize=”40″ titleclr=”#000000″ image_id=”2854″]

This is a PyTorch implementation of SoundStorm, a fast and parallel audio generation method from Google Deepmind. It uses MaskGiT to transform the residual vector quantized codes from Soundstream. The transformer model is based on Conformer, which is suitable for the audio domain.

[/vc_headings][vc_column_text]

Install

$ pip install soundstorm-pytorch

Usage

import torch
from soundstorm_pytorch import SoundStorm, ConformerWrapper

conformer = ConformerWrapper(
    codebook_size = 1024,
    num_quantizers = 4,
    conformer = dict(
        dim = 512,
        depth = 2
    ),
)

model = SoundStorm(
    conformer,
    steps = 18,          # 18 steps, as in original maskgit paper
    schedule = 'cosine'  # currently the best schedule is cosine
)

# get your pre-encoded codebook ids from the soundstream from a lot of raw audio

codes = torch.randint(0, 1024, (2, 1024))

# do the below in a loop for a ton of data

loss, _ = model(codes)
loss.backward()

# model can now generate in 18 steps. ~2 seconds sounds reasonable

generated = model.generate(1024, batch_size = 2) # (2, 1024)

To use raw audio as input, you have to feed your pre-trained SoundStream model to SoundStorm. You can train your own SoundStream model with audiolm-pytorch.

import torch
from soundstorm_pytorch import SoundStorm, ConformerWrapper, Conformer, SoundStream

conformer = ConformerWrapper(
    codebook_size = 1024,
    num_quantizers = 4,
    conformer = dict(
        dim = 512,
        depth = 2
    ),
)

soundstream = SoundStream(
    codebook_size = 1024,
    rq_num_quantizers = 4,
    attn_window_size = 128,
    attn_depth = 2
)

model = SoundStorm(
    conformer,
    soundstream = soundstream   # pass in the soundstream
)

# find as much audio you'd like the model to learn

audio = torch.randn(2, 10080)

# course it through the model and take a gazillion tiny steps

loss, _ = model(audio)
loss.backward()

# and now you can generate state-of-the-art speech

generated_audio = model.generate(seconds = 30, batch_size = 2)  # generate 30 seconds of audio (it will calculate the length in seconds based off the sampling frequency and cumulative downsamples in the soundstream passed in above)