A complete open-source system that runs general-purpose English speech-to-text and on-device neural text-to-speech entirely offline on ESP32-S3 and ESP32-P4 modules, with bindings for C, MicroPython and AtomVM (Erlang/Elixir). It ships a finetuned Citrinet-256 STT compressed to 4- or 8-bit (6.0 MB int4 or 9.8 MB int8 model files) plus a noisy-room variant, and tiny neural TTS voices (nano: 294k params, ~1.9 MB; heart: 2.27M params, additional ~2.4 MB for P4). STT needs about 6 MB flash and 84.5 KB internal RAM (plus a few hundred KB PSRAM during inference); TTS adds ~1.9 MB flash and ~0.5 MB working PSRAM. Models are memory-efficiently read in-place from flash (not copied into RAM). The ESP32-P4 runs faster (~0.30 s to transcribe 2 s of speech vs ~0.51 s on S3) and can host the larger heart voice; both boards are supported with drivers (ES7210/ES8311, PDM) and example hardware.
Model results and tooling are practical and specific: word-error-rate comparisons on LibriSpeech clean and 82 noisy-board recordings show tradeoffs (noisy-finetuned models reduce WER in noisy rooms at the cost of some clean-speech accuracy). The repo includes training/field-capture/datagen/finetuning pipelines, optimized 4/8-bit integer kernels for Xtensa/RISC-V vector units, build instructions (ESP-IDF v5.5.1, MicroPython v1.27.0, optional AtomVM), and examples for LoRa messaging, offline privacy-first voice control, and headless voice nodes. This targets projects needing local, private speech I/O rather than cloud accuracy or multilingual coverage.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
The Lightbulb Computer is a device that projects information into real-world spaces using modern projector technology and computer vision, avoiding traditional screens. It can be mounted in various locations and reacts to voice, gestures, and vision to enhance everyday tasks like cooking, studying, and planning. (lightbulbcomputer.com)
The switchdev model is an in-kernel driver framework for offloading switch device forwarding from the Linux kernel, enabling efficient data plane processing. It supports various setups, including data-center switches and SR-IOV, with components spanning user-space tools, kernel drivers, and hardware devices. (docs.kernel.org)
A RAM shortage has led to a resurgence of DDR4 memory, supported by new Intel and AMD CPUs compatible with older RAM, sockets, and chipsets. This comeback aims to address supply issues and meet ongoing demand for computer memory. (theverge.com)
The Museo de Historia de la Computación in Spain built a replica of the Cray-1 supercomputer using 30 obsolete Mac Minis. The project demonstrates how modern consumer hardware can emulate historic supercomputers and highlights the evolution of computing architecture. (arstechnica.com)
The Atari Falcon030 was Atari Corporation's final computer, released in late 1992 and based on a Motorola 68030 CPU with a digital signal processor. It was discontinued in 1993 as Atari shifted focus to the Jaguar console, but later versions were produced by C-Lab for music production. (atarimuseum.nl)
Global PC shipments have declined by 20 percent over the past year due to the RAMpocalypse. This drop reflects ongoing supply chain issues and market challenges in the computing industry. (techradar.com)
Today's best Hacker News stories, summarized and screenshotted, one email a day.