Spum Intelligence

A Neuroscape Imaging product. Published by Neuroscape Imaging Private Limited.

Spum brings local inference to the computer beside the instrument. This repository contains signed Mac releases, a native Ubuntu desktop/CLI release, and tested, openly licensed speech model packages. It does not contain proprietary engine source or private medical data.

Website: spum.si · Documentation · Speech guide · Whitepaper

Software releases

• CLI 0.2.4 for Apple silicon: includes the signed native speech runtime, notarized by Apple.
• Standalone Mac app 0.2.4: notarized app with the S menu-bar symbol and bundled CLI engine 0.2.4. Speech remains a CLI feature on macOS.
• Native Ubuntu desktop and CLI 0.2.4: x86_64, tested on Ubuntu 22.04. GTK window, CPU/CUDA inference, model import and English speech controls.
• Portable Linux x86_64 archive: keep its native runtime beside the engine.
• Ubuntu installation guide and functional release checks.
• Archive hashes: SHA256SUMS.txt.

curl -fsSL https://spum.si/install.sh | sh
# or
brew install vinayapathak/spum/spum

On Ubuntu, download the .deb above and run sudo apt install ./spum_0.2.4_amd64.deb, then open spum-desktop. Image input is currently macOS-only. There is no released or physically tested iPhone app.

Keep the included native runtime with the CLI, or use the installer/Homebrew package. No Python runtime is required.

Released speech models

ID	Task	Upstream
whisper-base-en	English speech recognition	OpenAI Whisper base.en, int8 ONNX conversion
kokoro-en-v1-int8	English speech synthesis	Kokoro 1.0 int8, three checked English voices

spum models catalog
spum models download whisper-base-en
spum models download kokoro-en-v1-int8
spum models verify whisper-base-en
spum transcribe recording.wav --json --output transcript.json
spum speak --text "The scanner is ready." --voice af_heart --output speech.wav

See models/ for model archives, their catalog, provenance, original licences and the small functional-check record. These are third-party pretrained models packaged for Spum; Neuroscape does not claim to have trained the upstream foundation models. Model files have separate licences described in LICENSES.md, not a blanket repository licence.

The recognition pack is English only. The synthesis interface exposes af_heart, af_alloy and am_adam, produces 24 kHz mono WAV, and checks pronunciation coverage before generating audio. Captions use recognition-window timing rather than word alignment. No microphone, remote speech API or voice-cloning facility is used.

Evaluation scope

Three reference recordings contained 77 normalized words with 3 word differences. Three synthesized phrases produced finite, non-silent audio, and an ASR roundtrip recovered their words. These small functional checks are not a general accuracy benchmark or an independent perceptual listening study. Spum is a research and engineering tool; these demonstrations do not establish clinical transcription or diagnostic validity.

Spum Intelligence 0.2.4

• Spum Reasoning: spum calculate provides deterministic numerical expressions, fluorescence SNR and packed microcamera-array throughput. The Mac app reviews model-extracted inputs and units before calculation, retains an audit record and renders equations with offline MathJax. Symbolic integration in the recorded DGX example uses an optional remote worker, not the local numerical parser.
• Remote execution: spum remote --host HOST --engine PATH -- COMMAND ARGS runs an existing engine over system SSH, requiring verified known hosts. Model paths refer to the selected remote host. This is private-network execution when you configure hosts and workers within that network, not inference on the Mac itself.
• Spum Voice: the native English recognition/synthesis packs remain available. A measured comparison with upstream Maya, Kokoro and eSpeak NG includes parameters, audio and ASR checks. Maya runs through a separately configured remote worker; its weights are not included here, and no new native Maya backend or Maya kernel speedup is claimed.
• Spum Afterglow: the name for the piano/MIDI continuation capability; the CLI remains spum music.

Spum Reasoning example · Voice benchmark and blind samples · Maths guide · SSH guide

Release checks: 62 numerical checks on macOS and Ubuntu, native Ubuntu package/GUI startup, H200 CUDA inference, Developer ID signing and Apple notarization. Checksums are supplied for the current release. The updated portrait and landscape films use actual app calculations; their original audio tracks are unchanged.