GigaType

private, unofficial fork of cjpais/Handy. gigatype preserves upstream mit attribution and adds gigaam multilingual plus verified windows cuda packaging. it is not endorsed by or affiliated with upstream handy.

gigatype is a local desktop speech-to-text application: press a shortcut, speak, and paste the transcription into the focused text field. version 0.9.3-gigatype.2 targets separate cpu and nvidia cuda 13 packages for windows x64.

what this fork adds

• four pinned gigaam multilingual ctc choices covering uz, kk, ky, ru, and en
• windows x64 cpu and nvidia cuda 13 packaging, with package-content and execution-provider audits
• explicit Auto, CPU, and CUDA behavior for onnx runtime models
• independent gigatype application identity, with upstream source history and mit attribution preserved
• no upstream updater or signing identity; private installers are unsigned and updates are installed manually

speech recognition, vad, history, and text insertion run locally. installers do not contain gigaam weights; selecting a model downloads its pinned files separately.

download: cpu or cuda

release 0.9.3-gigatype.2 targets the four windows x64 assets below. no packaged release is planned for other platforms.

edition	installer	use it when
cpu	GigaType_0.9.3-gigatype.2_x64-setup.exe	recommended for systems without a supported nvidia gpu
cpu	GigaType_0.9.3-gigatype.2_x64_en-US.msi	same cpu application in msi format
cuda 13	GigaType_0.9.3-gigatype.2_x64-cuda13-setup.exe	nvidia gpu package with app-local onnx runtime cuda 13 libraries
cuda 13	GigaType_0.9.3-gigatype.2_x64-cuda13_en-US.msi	same cuda application in msi format

choose one package format and one edition. the cpu edition is smaller and does not include cuda runtime libraries. the cuda edition requires a compatible nvidia display driver; cuda developer tools are not required.

when published, the four assets will be distributed with the private gigatype release. access requires permission to the private repository.

install on windows

1. after the private release is published, download one installer and find its sha256 in the matching release notes.
2. verify the local file with the command in verify sha256. stop if the values differ.
3. run the .exe setup or .msi; do not install both formats.
4. installers published by this private fork are unsigned, so windows may show an unknown-publisher or microsoft defender smartscreen warning. only after the sha256 matches, select More info -> Run anyway when that option is shown.
5. launch gigatype. on the windows onboarding microphone card, select Open System Settings. in Settings -> Privacy & security -> Microphone, enable Microphone access and Let desktop apps access your microphone, then return to gigatype. repeat this settings step until onboarding shows the permission as granted.

gigatype has no automatic updater. verify and install a newer private package manually when one is published.

first transcription

1. open gigatype and complete the windows microphone onboarding described above.
2. during onboarding, click Show all models before choosing a gigaam entry; gigaam models are not shown among the recommended cards. after onboarding, open Settings -> Models, where available models are listed directly without a Show all models control.
3. download and select GigaAM Multilingual 220M INT8 on cpu, or the matching gpu recommendation below. wait until the model is marked downloaded.
4. Auto is the default onnx runtime accelerator. to inspect or change it, open Settings -> Advanced, enable Experimental Features, then under Experimental use ONNX Acceleration. choose CPU or CUDA explicitly only when needed for the selected edition or diagnostics.
5. the default windows transcription shortcut is Ctrl+Space, with push-to-talk enabled. configure either setting if that workflow is not convenient.
6. focus a text field, hold the shortcut while speaking, then release it. if push-to-talk is disabled, press once to start and once to stop. gigatype transcribes locally and inserts the result into the focused application.

the first model selection needs network access because model weights are not embedded in the installer.

gigaam model guide

model	intended use	download size
GigaAM Multilingual 220M INT8	recommended cpu model	about 214 MiB plus vocabulary
GigaAM Multilingual 220M FP32 CUDA	recommended balanced gpu model	about 844 MiB plus vocabulary
GigaAM Multilingual 600M INT8	larger cpu model; use only when its extra memory and processing cost are acceptable	about 564 MiB plus vocabulary
GigaAM Multilingual 600M FP32 CUDA	recommended rtx 5080 model	about 2.18 GiB plus onnx graph and vocabulary

all four are multilingual ctc models for uz, kk, ky, ru, and en. 220M INT8 is the default recommendation for cpu use; 220M FP32 CUDA balances gpu download size and model capacity; 600M FP32 CUDA is the rtx 5080 recommendation.

rtx 5080 setup

1. install either cuda 13 asset, after verifying its sha256.
2. install a compatible current nvidia display driver. the gigatype package supplies onnx runtime, cuda 13.0 update 2 runtime components, and cudnn 9.16.0.29 beside the application; it does not supply the driver.
3. download and select GigaAM Multilingual 600M FP32 CUDA.
4. use Auto for normal operation. use explicit CUDA when you want startup/model-load failure instead of cpu fallback if cuda is unavailable.
5. inspect provider state from the gigatype installation directory when needed:

& .\GigaType.exe --list-accelerators --json

the repository verifier can measure cuda timing and exact-process vram use, but this readme does not promise a fixed speed or vram figure for every driver, audio sample, or machine.

accelerator behavior

gigaam and other onnx models use the onnx runtime (ORT) accelerator setting. whisper-family ggml/gguf models use the separate transcribe-cpp accelerator setting, so changing ort does not select a whisper backend.

• Auto: uses cuda only when this windows x64 build includes cuda support and the cuda execution provider registers successfully. otherwise it selects cpu and records the diagnostic reason.
• CPU: forces onnx inference onto cpu.
• CUDA: requires usable cuda support. missing app-local runtime files, provider registration failure, or an incompatible/missing nvidia driver causes a clear error; explicit cuda does not silently continue on cpu.

the cpu package is built without the cuda execution provider. the cuda package carries an audited app-local cuda dependency set. package verification checks that cpu installers contain no cuda runtime and cuda installers contain the required provider/runtime files.

model limitations

• supported languages are uz, kk, ky, ru, and en; other languages are outside this model bundle's contract.
• output is lowercase with little or no punctuation.
• the vocabulary has no digit tokens, so do not expect numeric digits in raw gigaam output.
• these ctc choices do not translate speech and do not stream partial text.
• optional post-processing can alter casing, punctuation, or number formatting, so compare raw output with post-processing disabled when evaluating the model itself.

privacy and model downloads

recording, vad, gigaam inference, history, and paste run on the local computer. gigaam itself does not send audio to a transcription service. optional cloud post-processing is a separate feature and can send transcript text to the provider you configure.

model weights download separately from pinned hugging face repositories. gigatype verifies every declared bundle file by expected size and sha256 before materializing it:

• istupakov/gigaam-multilingual-ctc-onnx at revision 458860e1983aef670dd9795fb6af603c82767d5d for both 220M choices
• istupakov/gigaam-multilingual-large-ctc-onnx at revision 07665ab5e54371dd1ac7b8b10f06478003723573 for both 600M choices

after a model is downloaded and verified, transcription can run without network access. deleting gigatype's app-data model directory requires downloading the files again; use the app's about/debug view to locate the exact app-data directory.

verify sha256

after publication, the authoritative hashes will be in the private release notes and will apply to the exact final asset names. for example:

$asset = Join-Path $env:USERPROFILE "Downloads\GigaType_0.9.3-gigatype.2_x64-cuda13-setup.exe"
(Get-FileHash -LiteralPath $asset -Algorithm SHA256).Hash.ToLowerInvariant()

compare all 64 hexadecimal characters with the value beside the same filename in the private release notes. filename similarity, file size, or a hash from another release is insufficient. if the value differs, delete the file and download it again; do not use Run anyway.

to inspect several downloaded assets:

Get-ChildItem (Join-Path $env:USERPROFILE "Downloads\GigaType_0.9.3-gigatype.2_x64*") -File |
  Get-FileHash -Algorithm SHA256 |
  Format-Table Path, Hash

troubleshooting

windows blocks the installer: verify sha256 first. when the hash matches the private release notes, use More info -> Run anyway; when it does not match, do not run the file.

cuda is unavailable: run & .\GigaType.exe --list-accelerators --json from the installation directory. confirm you installed a cuda13 asset and have a compatible nvidia driver. Auto may continue on cpu and report why; explicit CUDA returns an error. use the cpu package with GigaAM Multilingual 220M INT8 if gpu acceleration is not required.

model download fails: confirm the machine can reach hugging face, then retry from the model selector. proxy/firewall failures are reported against the failing bundle file; gigatype does not silently substitute another model. partial staging is cleaned before a later retry.

cpu transcription is too slow or memory-heavy: select GigaAM Multilingual 220M INT8, close other memory-heavy applications, and avoid an fp32 cuda-labeled model on cpu.

output lacks capitals, punctuation, or digits: this is expected raw gigaam behavior. use optional post-processing only if its privacy and provider behavior are acceptable.

a new private release is available: gigatype does not use upstream handy updates. download the new private asset and verify its release-specific sha256 before installing it.

build from source

see BUILD.md for exact prerequisites, clone commands, pinned runtime inputs, cpu/cuda package gates, and output paths. source development remains cross-platform, but 0.9.3-gigatype.2 targets packaged releases only for windows x64.

repository layout

• src/: react/typescript settings, onboarding, model selection, and overlay ui
• src-tauri/src/: rust application, audio pipeline, model catalog/bundles, transcription, shortcuts, and commands
• src-tauri/resources/: packaged vad resources and third-party notices; gigaam weights are deliberately absent
• scripts/: cross-platform helpers plus windows cpu/cuda build, audit, and verification entrypoints
• tests/ and src-tauri/tests/: frontend contracts and rust integration tests
• .github/workflows/: ci build and package-audit workflows
• docs/superpowers/: approved design and implementation-plan records for this fork

license and upstream attribution

gigatype preserves handy's git history and source attribution. the source remains under the mit license, including Copyright (c) 2025 CJ Pais. gigatype is an unofficial private fork with independent branding and no upstream endorsement.

model weights are separate works and are not relicensed by this repository's mit license. the multilingual onnx exports come from istupakov/gigaam-multilingual-ctc-onnx and istupakov/gigaam-multilingual-large-ctc-onnx, converted from ai-sage/GigaAM-Multilingual. review each model repository's model card and license before redistribution.

cpu packages include microsoft onnx runtime's license and third-party notices. cuda packages also include the official nvidia cuda/cudnn license files and THIRD_PARTY_NOTICES-CUDA.txt. those components remain under their respective licenses; nvidia driver libraries are supplied by the installed display driver and are not redistributed by gigatype.