v0.1.7Free and open source, forever

Speech tools that
never phone home.

Turn text into speech and speech into text, right on your computer. Whisper, Kokoro and Piper run on your own hardware. No account, no API key, no monthly bill.

Download TalkrRead every line of it

On another system? See every download

The Transcribe screen turning a 12 second interview into timestamped text with Whisper, offline, on the GPU.

Built on open models and runtimes you can inspect yourself

  • whisper.cpp
  • sherpa-onnx
  • Kokoro
  • Piper
  • Metal
  • Vulkan
  • Tauri

Features

Two tools. Zero servers.

One screen turns text into a voice. The other turns a voice into text. Everything in between stays on your disk.

Speak

Paste a paragraph, a script or a whole chapter. Pick a voice, set the speed, and save it as MP3, FLAC or WAV.

Transcribe

Record from the mic or drop in an MP3, WAV, FLAC, OGG or M4A. Get text back with timestamps.

Uses your GPU

Transcription runs on Vulkan or Metal, in its own process. No GPU, or it stops? Talkr quietly switches to the CPU.

History that keeps

Everything you make is saved locally with full-text search and favorites.

Take it anywhere

Export transcripts as text, Markdown, SRT or WebVTT subtitles, CSV or JSON, and speech as MP3, FLAC or WAV.

Model library

17 open models. One click each.

Download from inside the app, check the license first, delete whenever you like. Already have a GGML or ONNX file? Import it.

Speech to text

Whisper · GGML

  • Whisper TinyEnglish

    English · MIT

    78 MB

  • Whisper BaseEnglishRecommended

    English · MIT

    148 MB

  • Whisper SmallEnglish

    English · MIT

    488 MB

  • Whisper SmallEnglishCompressed · less memory

    English · MIT

    190 MB

  • Whisper BaseMultilingualRecommended

    10 languages · MIT

    148 MB

  • Whisper BaseMultilingualCompressed · less memory

    10 languages · MIT

    60 MB

  • Whisper SmallMultilingual

    10 languages · MIT

    488 MB

  • Whisper SmallMultilingualCompressed · less memory

    10 languages · MIT

    190 MB

  • Whisper MediumMultilingual

    10 languages · MIT

    1.5 GB

  • Whisper Large v3 TurboMultilingual

    10 languages · MIT

    1.6 GB

  • Whisper Large v3 TurboMultilingualCompressed · less memory

    10 languages · MIT

    574 MB

Text to speech

Kokoro, Piper · ONNX

  • Kokoro Multi-languageRecommended

    9 languages · Apache-2.0

    350 MB

  • Kokoro English

    English · Apache-2.0

    320 MB

  • Piper EnglishLessac, Medium

    English · MIT + GPL-3.0

    67 MB

  • Piper English UKAlan, Medium

    English (UK) · MIT + GPL-3.0

    67 MB

  • Piper GermanThorsten, Medium

    German · MIT + GPL-3.0

    67 MB

  • Piper FrenchSiwis, Medium

    French · MIT + GPL-3.0

    67 MB

Compared

Same job. Different address.

Cloud speech tools work well. They also need your audio on someone else's computer. Here is the trade, laid out plainly.

Talkr compared with a typical cloud speech service
FeatureTypical cloud serviceTalkr
Where your audio goesUploaded to their serversStays on your disk
Works offlineNoYes, after the model downloads
AccountRequiredNone
PricePer minute or monthlyFree, MIT licensed
SpeedDepends on your connectionDepends on your computer
Can you read the code?Usually notEvery line

Cloud tools still win on some things, like speaker labels and very long files on weak hardware. Talkr is for when the audio shouldn't leave the room.

Privacy

We couldn't read your files if we wanted to.

Audio, text and history live in ~/.talkr. Talkr only goes online when you download a model you picked. Delete the folder and it's all gone.

servers
0serversTalkr has no backend to send anything to.
accounts
0accountsNothing to sign up for, nothing to leak.
analytics
0analyticsNo tracking, no crash reporter, no pings.

Read the privacy policy

Open source

Don't trust us. Build it yourself.

The whole app is on GitHub under the MIT license: the Rust core, the interface, this website. Clone it, read it, change it, ship your own.

  • MIT

    License

  • Rust

    Core

  • Tauri 2

    Shell

View on GitHubReport an issue
Commands to build Talkr from source
~/code
# grab the source
$ git clone https://github.com/alihassan-coder/Talkr
$ cd talkr && pnpm install
# run the desktop app with hot reload
$ pnpm dev:desktop
✓ Rust core compiled · GPU backend: Metal
✓ Talkr is running. No servers were contacted.
# or just this website
$ pnpm dev:web

FAQ

Questions.
Short answers.

Something missing? Ask on GitHub

Is it really free?
Yes. Talkr is MIT licensed and has no paid tier. The models are open source too: Whisper is MIT, Kokoro is Apache 2.0, and Piper voices are MIT with espeak-ng data under GPL 3.0.
What does my computer need?
Windows 10+, macOS 11+, or Linux with glibc 2.35+ (Ubuntu 22.04 or newer), and 4 GB of RAM. The compressed Whisper models are made for smaller machines. On Intel and AMD, speech to text needs a processor with AVX2, which most have had since 2013. A GPU makes transcription much faster: Apple Silicon through Metal, or any Vulkan GPU from NVIDIA, AMD or Intel. Without one, Talkr uses the CPU. Text to speech always runs on the CPU, and it is quick there.
Does anything get uploaded?
No. Audio and text are processed on your machine. Talkr only goes online when you download a model you picked, and to check for updates at launch and every six hours. That check sends nothing about you, and you can turn it off in Settings.
Can it translate?
One way. Whisper can take speech in any language it supports and write it out in English. Turn on Translate before you transcribe.
Where are my files?
In ~/.talkr on macOS and Linux, and C:\Users\You\.talkr on Windows. Set TALKR_HOME to move it. Delete the folder and everything is gone.
Can I use my own models?
Yes. Use Import local model on the Models page. Speech to text takes a Whisper GGML .bin file. Text to speech takes a folder with model.onnx, tokens.txt and espeak-ng-data.

Your voice stays yours. Starting now.

Free for everyone. A 15 MB download. No sign-up screen, ever.

Download TalkrFork it on GitHub