Brethof Voice Pro — what's new

The running story of this app, one release at a time. If you are on the latest version, this is the list of what you already have — and the notes below the changelog explain the decisions behind some of it.

Release Aug 14, 2026

v2.3.1 — two-factor accounts can sign in again

A regression in the previous release left accounts with two-factor authentication locked out of signing in. v2.3.1 fixes it. If that was you: enter your credentials, enter your code, and you're in. Nothing else changed.

Release Aug 6, 2026

v2.2.2 — subtitle files, roughly twice as fast

Translating a subtitle file now takes about half the time it did. The timing is untouched — every timestamp lands exactly where it was in the original, whether you asked for sentence-level or word-level. If you subtitle for a living, this is the release where the wait got noticeably shorter.

Also in v2.2.2:

  • Microphone input fix — and it mattered most on good microphones. Audio levelling could push the loudest moments of a clean recording past full scale and clip them. The better your microphone, the wider the gap between its peaks and its average level, and the more likely this was to bite — a compressed, noisy input sailed through untouched while a quiet room and a decent condenser got clipped. Levelling now leaves headroom instead, so consonants and hard attacks survive intact.
  • New translate_batch MCP tool — translate many strings in one call, with a batch-size parameter that backs off automatically if VRAM runs short.
Download v2.2.2 →
Release Jul 11, 2026

v2.2.1 — the app you own stays yours

Brethof Voice Pro is yours once you buy it, and v2.2.1 makes sure nothing stands between you and it. The licence check at launch used to behave as a gate: no network, a busy server, a firewall that said no, and you could be kept out of software you had paid for. It now fails open — if the check can't reach us, the app opens anyway. Software you own should not ask permission to start. It doesn't anymore.

  • Re-check Licence button on the licence dialog — forces a fresh check and continues launching on success. Previously the only ways out were Buy or Quit.
  • No trial-expiry branding for paying customers — dialogs now say what actually happened instead of showing everyone the same message.
  • No more UI freezes — model downloads run off the interface thread with percentage progress, and startup licence checks no longer block the window on a slow connection.
  • No lost dictation — if a transcription is still running when you finish speaking, the new audio is queued rather than silently dropped, and a stalled worker can no longer type late text into your next recording.
  • Offline is quiet — no repeated network waits when there is no connection.
Download →
Release Jun 4, 2026

v2.2.0 — record the whole meeting, not the first five minutes

Long recordings now run as background jobs. Hand it a two-hour meeting file and go make coffee — there's no length limit and nothing to babysit. Leave it running and forget about it, and it finishes on its own and keeps the transcript for you. A three-hour call is no different from a three-minute one; it just takes longer.

  • Microphone or system audio — record your own voice, or loopback-capture whatever is playing on your machine: meeting calls, browser tabs, anything on your speakers. New list_loopback_sources tool to see what is available.
  • Nothing is thrown away — if a recording is abandoned, the watchdog now transcribes it to a retrievable job rather than discarding it.
  • Honest limits — the short instant-transcription mode was raised to a 10-minute ceiling, and asking for more than it can hold now returns a clear error pointing at the long-form tool. It no longer quietly clamps your audio and hands back a partial answer.
  • Linux translation fix — a missing dependency meant translation was broken in every Linux build since v1.7.0. Found by a new end-to-end release gate that drives the real, built binary before any release ships.
Download →
Update Jun 4, 2026

v2.1.0 — your agent gets an ear, over HTTP

An agent on the same machine can now use Brethof Voice Pro — transcribe, translate, record — through MCP. It works over local HTTP, so an agent that can only reach things by URL can use it without an API key and without anything leaving the machine. A token guards the door, and the server refuses to start without one. This one requires a paid licence.

  • Private by default — it listens only on your own machine unless you deliberately change that, and a wider bind refuses to start without an authentication token. When a token is set, every request must present it.
  • No interrupted work — models that unload when idle can no longer be unloaded in the middle of a long job. A timeout that ignores work in progress will eventually kill something legitimate, so the idle clock now pauses while anything is running.
Download →
Release May 21, 2026

v2.0.0 — translation that never leaves home

The big one. Translation now runs entirely on your machine, into 38 languages. Speak once and have it typed in English, Japanese, French and Spanish at the same time — one per line, or side by side in the same line. The engine is Tencent's Hunyuan MT2, in two sizes you choose: fast for everyday, quality for when it matters. At the quality size, Tencent's own published measurements put it at 97.9% of Google Gemini 3.1 Pro on the FLORES-200 benchmark, and ahead of it on real-world and minority-language translation. All of it offline. Nothing goes anywhere.

What is new in v2.0.0:

  • Transcribe + translate — the Transcribe popup has a "Translate to:" dropdown. ASR transcribes, MT translates, both render side-by-side. Works on files, mic, and system audio (loopback capture of meeting calls, browser tabs, anything playing on your speakers).
  • Multilingual voice keyboard — new translation chip on the main row opens a 3-column language grid. Pick one or more targets, the keyboard types the translation. Layout modes: one per line, inline (EN: … || PL: …), or first target only.
  • SRT / VTT subtitle translator — preserves timings + cue indices, supports bilingual mode (source line + translation per cue). Exposed via MCP for AI-agent automation.
  • Multi-GPU per engine — Settings → Models has an independent device picker for each engine. Run ASR on Vulkan 0 and translation on Vulkan 1, or pin Hunyuan MT2 7B to CPU on VRAM-tight laptops.
  • Personal voice training is fully live — LoRA fine-tuning on your own voice, end-to-end on your machine. The main window's training card shows total samples and minutes captured at a glance; "Start training" in the Training tab auto-picks NVIDIA CUDA or CPU, then auto-exports the trained model to GGUF. Free with every paid licence.
  • 4 new MCP toolstranslate_text, translate_srt, list_compute_devices, set_compute_device. Total tool count now 19.
  • Honest language splits — ASR: 30 selectable languages + auto-detect + 22 Chinese dialects (52 total). Translation: 38 languages (33 main + 5 Chinese-area minority). 23 work in both directions.

Linux binary is 161 MB, Windows installer is 118 MB. Existing licences carry over — just download v2.0.0 and the translation models will appear in Settings → Models.

Download v2.0.0 →
Release May 20, 2026

v1.6.16 — your voice becomes its own model

Brethof Voice Pro now trains on your voice. The training data is the corrections you already make — fix a misheard word once, and that becomes the example. You are never asked to read sample sentences. Training is one click, runs on your own hardware, and when it finishes the new model is ready to switch to.

The trial unlocks everything except personal training and the agent connection; this one is for people who use the app every day.

Engineering May 15, 2026

v1.6.0 — set it to work, come back later

Two things. Transcription can now run as an asynchronous job — drop the audio in, collect the transcript when it's done, don't sit in front of it.

And subtitle timing: you can now time word by word instead of sentence by sentence, for the moments when you need to edit precisely. If you make subtitles, you already know why that matters. ```

Engineering May 15, 2026

v1.5.0 — one runtime, and ONNX is gone

The ONNX runtime is gone. The whole product now runs on llama.cpp and our own translation library — one runtime, smaller on disk, faster to start, and easier to ship.

The models are single GGUF bundles, which is what makes the small one run on any 4 GB graphics card. There's a longer note on why we made the move further down the page.

Engineering May 14, 2026

v1.4.0 — your first agent connection, and a picker for your GPUs

Two firsts. The MCP server debuts here — the first way to give an agent ears, all on your own machine.

And the device picker: transcription and translation can now each choose their own processor, so one can run on your graphics card while the other works on the CPU. A laptop with a modest GPU can run both without grinding to a halt.

Roadmap Apr 16, 2026

macOS — in progress, Apple Silicon first

macOS is in development, Apple Silicon first. There is no date, and we are not going to give you one until the build is real. In the meantime, Linux and Windows are the same product with the same features — nothing is held back on one to sell the other. When the Mac build ships, this page will say so.

Read post →
Product Apr 16, 2026

Train it on your own voice

Your accent is not a problem to work around; it's a model to train. Brethof Voice Pro learns from the fixes you already make — correct a misheard word once, and that correction becomes an example. No sample sentences, no calibration script, no hour of reading into the mic. The pinned words you already set — brand names, jargon, colleagues' names — steer both the transcript and the translation, so a name is never localised by accident. Training is one click, runs on your own hardware, and the new model is ready to switch to when it finishes. The corrections you were already making were always teaching material. Now they train the model.

Read post →
Announcement Apr 16, 2026

Brethof Voice Pro is live

The first release. Talk instead of typing, into anything — your editor, your terminal, a chat box, a form — and the words land where the cursor is. Transcription runs entirely on your own machine, using Alibaba's Qwen3-ASR with DeepFilterNet3 for noise and Silero VAD for voice detection. All of it local. The call with your lawyer, the unreleased product, the patient notes — none of it leaves the computer it ran on, because there is nowhere for it to go. It works on a plane, on a train through a tunnel, in a hotel with bad wifi. Thirty languages, 22 Chinese regional dialects on their own. Runs on the laptop you already have — with a graphics card or without one; with one, any Vulkan 1.2+ GPU, NVIDIA, AMD or Intel Arc. On Linux it ships as a single portable file, needs glibc 2.38 or newer, and on Windows it's a normal installer that needs no administrator rights. A 14-day trial, no card — just an email.

Read post →
Licensing Apr 16, 2026

Your own licence now covers your own business

The individual licence now covers solo business and freelance work. If it's you — sole trader, one-person studio, contractor — the licence you buy covers your work. No per-seat accounting, no separate business tier. One licence, two devices, bought once. It comes with a year of updates included; after that year updates are optional, and the app keeps working on the version you have, indefinitely. Software you paid for stays yours.

Read post →
Product Apr 12, 2026

Thirty languages, and a UI that reads like a human wrote it

Brethof Voice Pro transcribes in 30 languages, recognises 22 Chinese regional dialects, and translates across 38. And the interface is fully translated into each of those 30 — translated by people who use the product, not poured through a machine translation and shipped. You can set the app to your language and it reads like it was written there, because it was.

Read post →
Engineering Apr 8, 2026

Why we moved from ONNX to GGUF

The first builds ran on ONNX Runtime. It worked, but it was heavy, it was awkward to ship, and it made the product bigger than it needed to be. v1.5.0 replaced it with a single runtime — llama.cpp for transcription, our own library for translation — and the models became single GGUF bundles. What you notice: smaller downloads, faster startup, and a product that fits more comfortably on a laptop. Two model sizes, a choice rather than a tier — the small one for laptops and integrated graphics, the larger for accented or noisy audio. You can switch at any time, and you don't need a graphics card at all if you'd rather not.

Read post →
Principles Apr 4, 2026

What "100% offline" means, precisely

Three things reach the network, and nothing else. A licence check when the app starts. An update check when the app starts. And the model downloads you start yourself. The licence check sends a licence key; the update check sends a version string. Neither sends audio, neither sends text, ever. Once the models are on disk and the update check is turned off, nothing goes out at all. That is not a marketing sentence — a firewall log will show you. The reason to run everything on your own machine is not that the cloud is a feature. It's that your voice is the thing you can't un-say: the call with your lawyer, the unreleased product, the patient notes. A tool that handles those should not need a network connection to do its job. It should work on a plane, and it should not be able to leak. That is the whole pitch, and it is a design decision, not a feature.

Read post →

Get updates in your inbox

One email per release, with the changelog above. No spam, no tips, no weekly newsletter. Unsubscribe in one click, whenever.

Create Account

Hear it when it ships

New releases, real benchmarks and the occasional deep-dive. No spam, unsubscribe in one click.

Everything we build

External:   YouTube · GitHub