4 ms·
So....I'm going to paste the abstract here because the headline is incredibly misleading and should be changed. >Abstract: We describe a neural network-based s
by grawprog 7y ago
So....I'm going to paste the abstract here because the headline is incredibly misleading and should be changed.
>Abstract: We describe a neural network-based system for text-to-speech (TTS) synthesis that is able to generate speech audio in the voice of many different speakers, including those unseen during training. Our system consists of three independently trained components: (1) a speaker encoder network, trained on a speaker verification task using an independent dataset of noisy speech from thousands of speakers without transcripts, to generate a fixed-dimensional embedding vector from seconds of reference speech from a target speaker; (2) a sequence-to-sequence synthesis network based on Tacotron 2, which generates a mel spectrogram from text, conditioned on the speaker embedding; (3) an auto-regressive WaveNet-based vocoder that converts the mel spectrogram into a sequence of time domain waveform samples. We demonstrate that the proposed model is able to transfer the knowledge of speaker variability learned by the discriminatively-trained speaker encoder to the new task, and is able to synthesize natural speech from speakers that were not seen during training. We quantify the importance of training the speaker encoder on a large and diverse speaker set in order to obtain the best generalization performance. Finally, we show that randomly sampled speaker embeddings can be used to synthesize speech in the voice of novel speakers dissimilar from those used in training, indicating that the model has learned a high quality speaker representation.
- aerovistae 7y agoDo you want to elaborate on how the title is misleading? From reading the abstract it seems accurate to me.
- corobo 7y ago"AI Clones Your Voice" implies there might be something on the linked page that involves an AI cloning my voice. Maybe a way to record a few phrases, maybe a text to speech that then uses my voice. Something like that. This does not do that - only provides pre-rendered samples, kinda disappointing. Impressive, but disappointing.
- lukeplato 7y agoI took the headline from this Two Minute Papers video: https://www.youtube.com/watch?v=0sR1rU3gLzQ https://www.youtube.com/watch?v=0sR1rU3gLzQ
- godelzilla 7y agoThanks... Too long of a scroll to find somebody posting the actual science behind the click-bait.
- azinman2 7y agoNit: this is more design/engineering than science. There is no hypothesis being tested about how the world works.
- umeshunni 7y agoDid you know that there's an entire field called Computer Science - https://en.wikipedia.org/wiki/Computer_science https://en.wikipedia.org/wiki/Computer_science
- amelius 7y agoComputer science should have been called "Computer math".
- JaRail 7y agoDid you read section 3 of the paper where they evaluate their system? > We primarily rely on crowdsourced Mean Opinion Score (MOS) evaluations based on subjective listening tests. All our MOS evaluations are aligned to the Absolute Category Rating scale [14], with rating scores from 1 to 5 in 0.5 point increments. We use this framework to evaluate synthesized speech along two dimensions: its naturalness and similarity to real speech from the target speaker. They're testing if the generated speech sounds natural with a well-defined and reproducible experiment. That's science.
- azinman2 7y agoEvaluation doesn’t make it science. There’s no investigation of the physical or natural world going on, unless they really think they’re modeling how humans are able to talk. But they’re not — they’re trying to create a system that works no matter how unnatural it is.