A Language Modeling Approach to Audio Generation

|paper| blog post|

Zalán Borsos, Raphaël Marinier, Damien Vincent, Eugene Kharitonov, Olivier Pietquin, Matt Sharifi, Dominik Roblek, Olivier Teboul, David Grangier, Marco Tagliasacchi, Neil Zeghidour

Google Research

Abstract. We introduce AudioLM, a framework for high-quality audio generation with long-term consistency. AudioLM maps the input audio to a sequence of discrete tokens and casts audio generation as a language modeling task in this representation space. We show how existing audio tokenizers provide different trade-offs between reconstruction quality and long-term structure, and we propose a hybrid tokenization scheme to achieve both objectives. Namely, we leverage the discretized activations of a masked language model pre-trained on audio to capture long-term structure and the discrete codes produced by a neural audio codec to achieve high-quality synthesis. By training on large corpora of raw audio waveforms, AudioLM learns to generate natural and coherent continuations given short prompts. When trained on speech, and without any transcript or annotation, AudioLM generates syntactically and semantically plausible speech continuations while also maintaining speaker identity and prosody for unseen speakers. Furthermore, we demonstrate how our approach extends beyond speech by generating coherent piano music continuations, despite being trained without any symbolic representation of music.

Speech continuation

Continuations using 3 second prompts from LibriSpeech test-{clean, other}, for speakers and content not seen during training. AudioLM excels at generating continuations that:

Librispeech test-clean
Original Prompt Continuations
Librispeech test-other
Original Prompt Continuations

Acoustic generation

For acoustic generation, we sample the acoustic tokens given the semantic tokens extracted from the original samples from LibriSpeech test-clean. The model generates samples with different speakers and recording conditions, while the semantic content is identical.

Original Acoustic generation (stage 2 and 3)

Unconditional generation

The unconditional generation performs sampling without using prompts. In that case, every sequence varies in speaker identity, linguistic content, and recording conditions.

Samples from unconditional generation

Generation without semantic tokens

To illustrate that the semantic tokens are crucial for generating coherent linguistic content, we train the language model on the acoustic tokens only. While the generated continuations of the 4-second prompts maintain speaker identity, the linguistic content is inconsistent, and often akin to babbling.

Continuations with a language model trained on the acoustic tokens only (without semantic tokens)

Comparing SoundStream reconstructions

We compare SoundStream reconstructions of two models, one using 3-layer residual vector quantization (3-RVQ) and another with 12 layers (12-RVQ), the latter being the default. The equivalent bitrates are 1.5 kbps and 6 kbps.

Original SoundStream 3-RVQ SoundStream 12-RVQ

Piano continuation

AudioLM is not limited to modeling speech. It can also learn to generate coherent piano music continuations, despite being trained on piano music without any symbolic representation. We also show the continuations produced by a version of the model trained exclusively on the acoustic tokens. These continuations are much less coherent, stressing the importance of the semantic tokens in our framework. The 4-second prompts come from the test split of MAESTRO dataset.

Original Prompt Continuation by acoustic-only model Continuation by AudioLM