Dictave
Resources · Research
The research we stand on
A map of the field, not a parts list. We deliberately do not say which research is inside Dictave — discretion about suppliers is part of the privacy posture — but the science deserves pointing at. Every link below is the work of the lab that published it, opens on their site, and carries none of your browsing context with it.
Speech to text
-
Sequence Transduction with Recurrent Neural Networks
The idea behind recognisers that type while you are still talking, rather than waiting for you to finish.
-
Deep Speech 2: End-to-End Speech Recognition in English and Mandarin
An early proof that one network, given enough speech, could replace an entire pipeline of hand-built parts.
-
wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations
How machines learn the structure of speech from audio nobody transcribed.
-
Conformer: Convolution-augmented Transformer for Speech Recognition
The architecture that quietly sits underneath much of modern recognition.
-
Robust Speech Recognition via Large-Scale Weak Supervision
The Whisper paper: what 680,000 hours of found audio buys you in robustness across accents and noise.
-
Fast Conformer with Linearly Scalable Attention
Recognition made cheap enough to run well away from a data centre — the line of work that makes local dictation plausible at all.
-
Scaling Speech Technology to 1,000+ Languages
The massively multilingual push: most of the world's languages have almost no transcribed audio.
-
Google USM: Scaling Automatic Speech Recognition Beyond 100 Languages
The same ambition from the other giant, with different trade-offs worth comparing.
Text to speech — the other direction
Dictave listens; it does not talk back. But machines that speak are the other half of this field, the research communities overlap almost entirely, and anyone curious about one should read the other.
-
WaveNet: A Generative Model for Raw Audio
The paper that made machines sound human — one audio sample at a time, absurdly slowly, and it changed everything.
-
Natural TTS Synthesis by Conditioning WaveNet on Mel Spectrogram Predictions (Tacotron 2)
Text in, natural speech out, end to end.
-
FastSpeech 2: Fast and High-Quality End-to-End Text to Speech
Making synthesis fast enough to be useful rather than merely impressive.
-
Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech (VITS)
A single model that learns to speak with rhythm and variation instead of a flat reading voice.
-
Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers (VALL-E)
Three seconds of a voice, imitated. Impressive and unsettling in equal measure — read it and you will understand why voice privacy matters.
The field's honest problems
Speech recognition is not solved, and pretending otherwise is how bad products get shipped. Two papers everyone building or buying dictation should read:
-
Careless Whisper: Speech-to-Text Hallucination Harms
Recognisers can confidently invent words nobody said. This is why accuracy has to be a measured gate, not a marketing adjective.
-
Racial disparities in automated speech recognition
Recognisers work measurably worse for some voices than others. The honest response is to measure against diverse speech and publish nothing you have not measured.