Applied sciences
Our pioneering speech era applied sciences are serving to folks around the globe work together with extra pure, conversational and intuitive digital assistants and AI instruments.
Speech is central to human connection. It helps folks around the globe change data and concepts, specific feelings and create mutual understanding. As our expertise constructed for producing pure, dynamic voices continues to enhance, we’re unlocking richer, extra participating digital experiences.
Over the previous few years, we’ve been pushing the frontiers of audio era, creating fashions that may create top quality, pure speech from a spread of inputs, like textual content, tempo controls and explicit voices. This expertise powers single-speaker audio in lots of Google merchandise and experiments — together with Gemini Dwell, Challenge Astra, Journey Voices and YouTube’s auto dubbing — and helps folks around the globe work together with extra pure, conversational and intuitive digital assistants and AI instruments.
Working along with companions throughout Google, we just lately helped develop two new options that may generate long-form, multi-speaker dialogue for making advanced content material extra accessible:
- NotebookLM Audio Overviews turns uploaded paperwork into participating and energetic dialogue. With one click on, two AI hosts summarize person materials, make connections between subjects and banter forwards and backwards.
- Illuminate creates formal AI-generated discussions about analysis papers to assist make information extra accessible and digestible.
Right here, we offer an summary of our newest speech era analysis underpinning all of those merchandise and experimental instruments.
Pioneering strategies for audio era
For years, we have been investing in audio era analysis and exploring new methods for producing extra pure dialogue in our merchandise and experimental instruments. In our earlier analysis on SoundStorm, we first demonstrated the flexibility to generate 30-second segments of pure dialogue between a number of audio system.
This prolonged our earlier work, SoundStream and AudioLM, which allowed us to use many text-based language modeling strategies to the issue of audio era.
SoundStream is a neural audio codec that effectively compresses and decompresses an audio enter, with out compromising its high quality. As a part of the coaching course of, SoundStream learns the way to map audio to a spread of acoustic tokens. These tokens seize all the data wanted to reconstruct the audio with excessive constancy, together with properties equivalent to prosody and timbre.
AudioLM treats audio era as a language modeling activity to supply the acoustic tokens of codecs like SoundStream. In consequence, the AudioLM framework makes no assumptions concerning the kind or make-up of the audio being generated, and may flexibly deal with a wide range of sounds with no need architectural changes — making it a very good candidate for modeling multi-speaker dialogues.
Instance of a multi-speaker dialogue generated by NotebookLM Audio Overview, primarily based on a couple of potato-related paperwork.
Constructing upon this analysis, our newest speech era expertise can produce 2 minutes of dialogue, with improved naturalness, speaker consistency and acoustic high quality, when given a script of dialogue and speaker flip markers. The mannequin additionally performs this activity in below 3 seconds on a single Tensor Processing Unit (TPU) v5e chip, in a single inference go. This implies it generates audio over 40-times quicker than actual time.
Scaling our audio era fashions
Scaling our single-speaker era fashions to multi-speaker fashions then turned a matter of knowledge and mannequin capability. To assist our newest speech era mannequin produce longer speech segments, we created an much more environment friendly speech codec for compressing audio right into a sequence of tokens, in as little as 600 bits per second, with out compromising the standard of its output.
The tokens produced by our codec have a hierarchical construction and are grouped by time frames. The primary tokens inside a gaggle seize phonetic and prosodic data, whereas the final tokens encode tremendous acoustic particulars.
Even with our new speech codec, producing a 2-minute dialogue requires producing over 5000 tokens. To mannequin these lengthy sequences, we developed a specialised Transformer structure that may effectively deal with hierarchies of data, matching the construction of our acoustic tokens.
With this system, we will effectively generate acoustic tokens that correspond to the dialogue, inside a single autoregressive inference go. As soon as generated, these tokens might be decoded again into an audio waveform utilizing our speech codec.
Animation displaying how our speech era mannequin produces a stream of audio tokens autoregressively, that are decoded again to a waveform consisting of a two-speaker dialogue.
To show our mannequin the way to generate reasonable exchanges between a number of audio system, we pretrained it on a whole bunch of 1000’s of hours of speech information. Then we finetuned it on a a lot smaller dataset of dialogue with excessive acoustic high quality and exact speaker annotations, consisting of unscripted conversations from quite a lot of voice actors and reasonable disfluencies — the “umm”s and “aah”s of actual dialog. This step taught the mannequin the way to reliably swap between audio system throughout a generated dialogue and to output solely studio high quality audio with reasonable pauses, tone and timing.
In keeping with our AI Rules and our dedication to creating and deploying AI applied sciences responsibly, we’re incorporating our SynthID expertise to watermark non-transient AI-generated audio content material from these fashions, to assist safeguard towards the potential misuse of this expertise.
New speech experiences forward
We’re now targeted on bettering our mannequin’s expressivity, acoustic high quality and including extra fine-grained controls for options, like prosody, whereas exploring how greatest to mix these advances with different modalities, equivalent to video.
The potential functions for superior speech era are huge, particularly when mixed with our Gemini household of fashions. From enhancing studying experiences to creating content material extra universally accessible, we’re excited to proceed pushing the boundaries of what’s doable with voice-based applied sciences.
