Parakeet Redux Makes Local ASR Tiny

// 1h agoMODEL RELEASE

Parakeet Redux Makes Local ASR Tiny

Moondream’s Parakeet Redux compresses NVIDIA’s 0.6B Parakeet model into a 178MB, 1.58-bit speech recognizer for fast CPU and Apple Silicon inference. It supports 25 languages and preserves near-original English accuracy, though noisy audio remains a clear weakness.

// ANALYSIS

Parakeet Redux makes local transcription genuinely practical on laptops and edge devices, proving aggressive quantization can deliver speed without completely sacrificing quality.

–Ternary weights shrink the model from 1.2GB to 178MB and enable 113× real-time throughput on eight CPU cores.
–English WER rises only slightly, from 6.26% to 6.55%, while multilingual FLEURS results improve over the original.
–Background noise is the key tradeoff: Redux reaches 9.04% WER versus 6.72% for the base model.
–Photon’s hardware-specific kernels make the model especially compelling for CPU-first, privacy-sensitive transcription applications.

// TAGS

parakeet-reduxmoondreamspeechsttquantizationedge-aiinference

DISCOVERED

1h ago

2026-10-04

PUBLISHED

1h ago

2026-10-04

RELEVANCE

10/ 10

AUTHOR

Better Stack

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

RESEARCH21m ago

Finetuning with Sampling Makes SFT Rival RL

Harvard researchers introduce an MCMC-based projection-sampling method that reshapes expert off-policy traces toward a model’s own distribution before supervised fine-tuning. Across chemistry, math, and medical QA tasks, the approach reportedly improves generalization and reduces forgetting versus strong RL and self-distillation baselines.

RESEARCH21m ago

Sharpening Tax Exposes RL’s Coverage Trade-Off

This research finds that RL post-training can improve pass@1 while narrowing an LLM’s solution coverage under repeated sampling. It introduces Sharpening Tax to measure that loss and PTGS, which adapts training temperature to prompt difficulty.

OPEN SOURCE53m ago

Crawlers Debuts Open Crawler Sandbox

Crawlers is an open-source, MIT-licensed browser sandbox for configuring autonomous crawler swarms, running them through a deterministic synthetic web, and inspecting or forking each simulation locally.