A transcriber that knows when you stopped.
Dictate pairs a compact speech co-processor with the connectome. The co-processor produces words; the brain decides segmentation, turn-taking and when a command is complete.
Transcription is a solved problem for classical models. What they are bad at is the timing: when to stop listening, when a pause is a full stop. Dictate leaves the words to the co-processor and gives the timing to a brain that has done turn-taking for a hundred million years.
Endpoint detection 3x closer to human.
Dictate ends a segment within 180 ms of where a human annotator would, against 550 ms for a fixed-silence threshold, at a word error rate matching its co-processor.
target Figures marked "target" are design targets for the first release, not measurements.
Same brain, its own wiring.
Input adapters occupy afferent regions, output adapters read efferent ones. Co-processors feed in through the embedding port.
- FlyCore v0.1.0
- weights v0.1.0-beta
- audio (antennae)
- speech embedding
- segments (text, t0, t1)
- speech recogniser
import flycore
brain = flycore.load("flylabs/dictate") # engine + weights + adapters
for seg in brain.stream(flycore.io.Microphone()):
print(seg.text, seg.t0, seg.t1)
# "open the settings" 3.10 4.02Words, with the brain on timing.
Voice commands for devices
Know the command is finished without a wake word or a button.
Meeting notes on device
Segments that match how people actually speak.
- Beta: co-processor covers 25 languages; the brain is language-agnostic.
- Word accuracy is the co-processor's, not the brain's.