I announced the voice conversational model named "Dessa" on 2026‑10‑25.
The model is associated with the company or project "Block".
I highlighted the launch with the hashtags #voiceturing and #aiaiaooo, signalling a goal of human‑level voice interaction and targeting the AI‑enthusiast community.
On 2026‑10‑24 I announced the open‑source Phonon‑2.5 multilingual speech model, which shrinks the original 2.5 GB size to about 6.2 % (≈ 155 MB) while achieving 101.8 % of the baseline word‑accuracy across 25 languages. This trade‑off is meant to enable on‑device dictation in the Detta app on macOS.
Phonon suite total parameter count is under 50 million.
Phonon‑31m model has 31 million parameters and is released under the Apache 2.0 license.
I tested NeuDecide, our speech‑to‑tool lightweight library, and the full session—including audio waveform, transcript, and tool‑call payload—was serialized into a 43 MB package, showing a modest storage footprint for debugging and audit trails.
NeuDecide.capture_size = 43 MB
NeuDecide.capture_timestamp = 2026-10-20 07:00 UTC
Bulbul v4 flash event was logged (event = Bulbul v4 flash).
The event timestamp is 2026-10-19 11:15 UTC (timestamp = 2026-10-19 11:15 UTC).
I praised the Phonon‑2 speech model for its tiny 164 MB download size and its ability to transcribe an hour of audio in just 20 seconds on a MacBook Air, positioning it as a more accurate per‑byte alternative to OpenAI’s Whisper large, which is roughly ten times larger.