JVM runs NVIDIA Parakeet ASR at up to 2X the C++ reference speed, no GPU or Python needed
mkheck · x · 2026-09-24
The open-source jinfer-parakeet project implements NVIDIA Parakeet speech recognition entirely on the JVM — no GPU, Python, ONNX Runtime, or whisper.cpp required. It claims up to 2X the speed of the C++ reference implementation on ordinary CPUs, transcribing an hour of audio in 44 seconds.
- At F16 the transcript matches parakeet.cpp word-for-word; remaining differences are quantization noise
- Streaming support with a draft tail that keeps up with the speaker, plus word timings and per-token confidence
- Parakeet v3 handles multilingual transcription with automatic punctuation, no language flag needed
- First-class GraalVM Native Image support for fast startup and low memory
A practical option for teams wanting local, CPU-only ASR without leaving the Java ecosystem.
Related event: Parakeet.java Runs NVIDIA ASR on Pure JVM, Twice as Fast as C++(2 posts)→
More from Infra
- Magnitude: open-source tool profiles your PC and auto-tunes the best local model for agents — anselm · 2026-09-24
- Data-center spending now tops canals, railroads and grid combined at 3.63% of GDP — kimmonismus · 2026-09-24
- AI data-center spending tops railroads, highways and grid, averaging 3.63% of GDP — kimmonismus · 2026-09-24
- Silicon Data pushes back: falling token index means cheaper mix, not stalled AI demand — AccBalanced · 2026-09-24
- Dev Runs Miles Training on H100s via Coding Agent, Ships First PR — VoidAsuka · 2026-09-24
- Industry responds to hyperscale RDMA paper with Multipath Reliable Connection on path to Ultra Ethernet — thoefler · 2026-09-24