Sub-1B speech-to-text model for on-device dictation, claiming better accuracy than rivals
The Hook
“Where is the Mona Lisa painting located today?”
Feature Teaser
“Is scarlet fever lethal?”
Feature Teaser
“Who won the first Nobel Prize for Poetry?”
Feature Teaser
“How big is Ohio?”
Feature Teaser
“What was Shakespeare's most famous poems?”
Feature Teaser
“Who colonized Spain?”
Feature Teaser
“How many autoimmune diseases are there?”
Feature Teaser
“Is vanilla native to Madagascar?”
Problem Agitation
“Who is head of Marvel Studios?”
Call to Action
“What are the three most consumed cereal grains?”
Synthesizable in Remotion · key components of the speech-model-benchmark-card template.
Dynamic audio waveform bars scrolling horizontally in grey, with active blue progress fill tracking playback time.
Render an array of vertical rounded rects whose heights sample amplitude arrays, overlaid with an interpolation mask driven by frame / fps.
Words render progressively; recognized words take subtle highlighted pills with red border outlines highlighting model hallucinations or transcription errors.
Map transcription timestamps to words. Wrap divergent tokens in rounded border spans styled with rgba(239, 68, 68, 0.4) based on error diff arrays.
Purpose. Demonstrate robustness across global accents and acoustic interference.
Execution. Rapidly cycling badge headers showing country flags and noise emojis above comparative outputs.
Purpose. Expose competitor model failures under background speaker noise.
Execution. Outlining incorrect competitor tokens in crimson red boxes while target model remains accurate.