Source-available • BSL 1.1

Developer dictation that learns your vocabulary locally.

TalaX turns Whisper into a dictation workflow for technical work: hotkey in, speak naturally, correct once, and reuse those corrections automatically the next time you talk.

Built for technical terms Fix “cuber netties” once and TalaX stops fighting your tooling vocabulary.

Local audio path Transcription and correction stay on the machine. No cloud audio, no subscription loop.

Correction memory Three layers learn from your edits instead of discarding them like generic dictation tools.

Context-aware profiles Keep separate vocabularies for DevOps, research, or personal writing without retraining from scratch.

Install TalaX View source

Current install path: build from source. No binary releases are published yet.

Live correction

Raw Whisper output is refined before it reaches your active app.

local only

Whisper

"deploy the cuber netties pod to staging"

Corrected output

"deploy the Kubernetes pod to staging"

Learning loop

After repeated confident fixes, TalaX applies the pattern for you.

3 layers

Dictionary corrector applies learned replacements first.

N-gram model scores context to catch improbable words.

Heuristics recover accents, compounds, acronyms, and similar sounds.

161 engine tests Coverage across audio, hotkeys, database, correction logic, and integration paths.

<61 ms pipeline latency Fast enough to feel like dictation, not a detached batch transcription job.

0 bytes cloud audio No remote transcription hop in the product story or in the installed workflow.

CPU no GPU required whisper.cpp keeps the path local and practical on normal developer hardware.

How it works

Three steps. One habit. Better dictated code and prose.

TalaX should feel like a native keyboard shortcut, not a separate transcription ritual. Speak, release, review, and let the corrections compound.

Hold the hotkey

Trigger dictation from any app with a push-to-talk shortcut instead of switching into a browser or cloud recorder.

Speak naturally

Whisper captures the raw transcript locally while TalaX keeps the context of technical vocabulary close by.

Correct once; reuse later

Fix the bad term, keep typing, and let the 3-layer pipeline remember how you actually talk about your stack.

Why existing dictation breaks

Generic speech tools were not made for developer language.

Technical vocabulary gets mangled. “Kubernetes”, “Terraform”, “Postgres”, or hostnames get turned into natural-language guesses.
Accents get punished twice. Lower baseline accuracy becomes a permanent problem when corrections are never reused.
Cloud capture changes the trust model. Voice, architecture conversations, and code context leave the machine.
Subscriptions do not buy personalization. You keep paying, but the tool still behaves like a stranger to your domain.

Learning, not just transcription

Corrections feed a persistent pipeline instead of vanishing at the end of the session.

Profiles for different contexts

Keep separate vocabularies for DevOps, research, and personal writing without cross-polluting terms.

Local-first architecture

Whisper, SQLite, and the correction engine stay on-device, so privacy is not a marketing afterthought.

Fast native stack

Tauri + Rust keep the app lightweight while the correction pipeline stays interactive.

Before / after proof

Show the mistake. Show the fix. Show the memory.

This is the product moment the page should prove immediately: raw speech gets corrected into something a developer would actually send to an editor or terminal.

Before TalaX

Raw transcript injected as-is.

friction

12kubectl rollout status deploy/cuber-netties-api

13terraform apply -target=aws_security_group_rule.allow-post grass

14open incident for fox dash and zero

15# you stop, fix terms manually, and break flow

With TalaX

Corrected output after the pipeline refines the phrase.

learned

12kubectl rollout status deploy/kubernetes-api

13terraform apply -target=aws_security_group_rule.allow-postgres

14open incident for fox-n0

15Pattern learned: "cuber netties" → "Kubernetes"

Context-aware profiles

Different projects need different vocabularies.

A DevOps profile should learn cluster names and Terraform resources. A research profile should remember model families and tooling terms. TalaX keeps those correction histories separate so you do not have to retrain from zero every time your domain changes.

work-devops research-ml personal clone / reset / switch SQLite learning store serialized n-gram state

Technical credibility

Native stack. Fast path. Clear correction model.

Technical depth matters here, but it comes after the user outcome. These are the pieces that make TalaX credible once the value is obvious.

Whisper via whisper-rs

Local speech-to-text with model sizes from ~75 MB to ~574 MB, downloaded on first use through the model manager.

3-layer correction pipeline

Dictionary replacement, trigram context scoring, and heuristic recovery for accents, acronyms, compounds, and fuzzy matches.

SQLite + profile state

Persistent correction history, sessions, and per-profile learning data without introducing a remote dependency.

Tauri + Rust app shell

Desktop integration for global hotkeys, audio capture, text injection, tray state, and a lightweight frontend.

Source-available licensing

Read the code. Verify the local path. Ship with eyes open.

TalaX is released under Business Source License 1.1 and converts to Apache 2.0 on 2030-03-28. The site uses the same wording everywhere so the trust model stays clear.

BSL 1.1 Apache 2.0 in 2030 161 tests 0 cloud dependencies in product flow

Ready to evaluate it?

Install TalaX and test it against your real vocabulary.

The current path is source build, with documented prerequisites and model choices on the install page.

Open install guide Browse GitHub