audiolm-google-torch
AudioLM: a Language Modeling Approach to Audio Generation
AudioLM is a model that generates audio in the waveform domain.
It uses 2 tokenizers: Soundstream to compute the Acoustic tokens and w2v-BERT to compute the Semantic tokens.
Soundstream: Acoustic Tokens
Soundstream [2] is a SOTA neaural audio codec. The model has 3 parts:
- Encoder
- Residual Vector Quantizer (RVQ)
- Decoder
The convolutional encoder/decoder takes a single channel waveform $x \in R^T$ and reconstructs it $\hat{x} \in R^T$ from the quantized embeddings. The embeddings are discretized using a residual vector quantizer (RVQ) with $Q$ vector quantizers each one composed by $N$ vocabulary symbols.
- Input: waveform at 16kHz.
- Encoder Embeddings: 50Hz (x320 reduction).
- Codebook symbols: $Y \in { 1, . . . , N }^{ T_A \times Q}$ where $T_A = T/320$
- Encoded audio: $enc(x) = R^{S \times D}$.
- One-hot encoded vectors shape: $S \times D$
- Decoder Embeddings.
- Reconstructed waveform.
Parameters:
- Encoder conv. blocks: $B_{enc} = 4$
- Decoder conv. blocks: $B_{dec} = 4$
- Channels: $C_{enc} = C_{dec}$
- Input samples: $M$
- One embedding: $M = 2\cdot 4\cdot 5\cdot 8 = 320$ (4 encoder strides: 2, 4, 5, 8).
- Embeddings dimensionality: $D$
- Number of samples in the time domain: $T$
- Number of samples of the encoded audio: $S = T / M$
Discriminator
The adversarial loss is computed with a STFT-based discriminator. The input to the STFTDiscriminator is the complexvalued STFT of the input waveform (real and imaginary parts) and the output, the logits.
w2v-BERT: Semantic Tokens
w2v-BERT [3] is a Transformer-based model for learning self-supervised audio representations. It maps a input waveform to a set of linguistic features.
Training
Random sampling $ [1:N_q] $ $ \rightarrow $ $ Q_i \quad i=1…n_q $ quantizers
Inference
Set $ n_q $ to change the desired bitrate value.
References
[1] AudioLM: a Language Modeling Approach to Audio Generation