AudioPaLM

A Large Language Model That Can Speak and Listen

|paper|

Paul Rubenstein*, Chulayuth Asawaroengchai*, Duc Dung Nguyen*,
Ankur Bapna, Zalán Borsos, Félix de Chaumont Quitry, Peter Chen, Dalia El Badawy, Wei Han, Eugene Kharitonov, Hannah Muckenhirn, Dirk Padfield, James Qin, Danny Rozenberg, Tara Sainath, Johan Schalkwyk, Matt Sharifi, Michelle Tadmor Ramanovich, Marco Tagliasacchi, Alexandru Tudor, Mihajlo Velimirović, Damien Vincent, Jiahui Yu, Yongqiang Wang, Vicky Zayats, Neil Zeghidour, Yu Zhang, Zhishuai Zhang, Lukas Zilka, Christian Frank

Google

Abstract. We introduce AudioPaLM, a large language model for speech understanding and generation. AudioPaLM fuses text-based and speech-based language models, PaLM-2 [Anil et al., 2023] and AudioLM [Borsos et al., 2022], into a unified multimodal architecture that can process and generate text and speech with applications including speech recognition and speech-to-speech translation. AudioPaLM inherits the capability to preserve paralinguistic information such as speaker identity and intonation from AudioLM and the linguistic knowledge present only in text large language models such as PaLM-2. We demonstrate that initializing AudioPaLM with the weights of a text-only large language model improves speech processing, successfully leveraging the larger quantity of text training data used in pretraining to assist with the speech tasks. The resulting model significantly outperforms existing systems for speech translation tasks and has the ability to perform zero-shot speech-to-text translation for many languages for which input/target language combinations were not seen in training. AudioPaLM also demonstrates features of audio language models, such as transferring a voice across languages based on a short spoken prompt.

AudioPaLM Overview

The AudioPaLM model, illustrated on speech-to-speech translation and automatic speech recognition. We take a pretrained text-only model (dashed lines) and expand its embeddings matrix to model a new set of audio tokens. The model architecture is otherwise unchanged; a mixed sequence of text and audio tokens is fed as input and the model decodes text or audio tokens. Audio tokens are converted back to raw audio with latter AudioLM stages.

Table of Contents

Speech-to-speech translation

In this section, we demonstrate AudioPaLM's ability to preserve the original speaker voice even in the translated audio. There are 3 groups of languages from CVSS-T dataset, as in Table 3 of Jia et al., 2022. For each language, we selected 5 utterances uniformly at random from the utterances which were sent for subjective ratings. The selection of the utterances for subjective evaluation is described in the AudioPaLM paper.

The columns are as follows:

  1. Original audio in the CVSS-T example
  2. The audio in the target language, as stored in the CVSS-T example
  3. The English accented audio in the target language, as output by AudioPaLM for the given original audio
  4. The source-language accented audio in the target language, as output by AudioPaLM for the given original audio
  5. The audio in the target language, as output by Translatotron 2 (Jia et al., 2022) for the given original audio, without voice preservation

High-resource languages

Original CVSS-T (ground truth target) AudioPaLM translation with English accent AudioPaLM translation with the source-language accent Translatotron 2 (prior work)

Medium-resource languages

Original CVSS-T (ground truth target) AudioPaLM translation with English accent AudioPaLM translation with the source-language accent Translatotron 2 (prior work)

Low-resource languages

Original CVSS-T (ground truth target) AudioPaLM translation with English accent AudioPaLM translation with the source-language accent Translatotron 2 (prior work)

Speech-to-text translation

Here we show English translation of the original audio, as translated by AudioPaLM. Note that usually there are multiple correct ways to translate a sentence, so a correct translation does not need to exactly match the targets from the CVSS-T dataset. The output of AudioPaLM currently does not have punctuation, because the training data did not have it. However, in our future work we plan to also add the punctuation.

High-resource languages

Source language audio CVSS-T target language transcript Audio PaLM translation

Medium-resource languages

Source language audio CVSS-T target language transcript Audio PaLM translation

Low-resource languages

Source language audio CVSS-T target language transcript Audio PaLM translation

ASR

In this section we show a few examples, per language, of the AudioPaLM transcribing the original audio.

High-resource languages

Original CVSS-T original transcript Audio PaLM transcription

Medium-resource languages

Original CVSS-T original transcript Audio PaLM transcription

Low-resource languages

Original CVSS-T original transcript Audio PaLM transcription