Typd’s dictation model learned from speech that other people recorded and shared under open licences, and the app runs on open-source software. This page names them. The datasets were used only to train the model; none of them ships with Typd or is shared again.
The base model
Both of Typd’s models are fine-tuned versions of Qwen3-ASR, made by the Qwen team at Alibaba Cloud and licensed under the Apache License 2.0. The full model starts from Qwen3-ASR-1.7B, the light model from Qwen3-ASR-0.6B. The light model also learned from a fine-tuned 1.7B model, which acted as its teacher.
The audio encoder file that comes with each model is ggml-org’s GGUF conversion of Qwen3-ASR’s encoder (1.7B, 0.6B), used unchanged. The notice and full licence text are on the model licence page.
Speech the models learned from
For each dataset: who made it, its licence, which model learned from it, and what I did with it. “Full” is the 1.7B model; “light” is the 0.6B model for smaller computers.
- Emilia-YODAS (English), by the Amphion team (He et al., 2025). CC BY 4.0. Full model. Talk from YouTube, cut into 5 to 90 second utterances, transcribed, and given finished text. Only the Emilia-YODAS part was used, never the non-commercial Emilia part.
- YODAS2, YODAS-Granary and YODAS, by ESPnet (Li et al., 2024), with Granary transcripts by NVIDIA (Koluguri et al., 2025). The audio comes from YouTube videos their creators published under Creative Commons. CC BY 3.0. Full and light models. English clips, segmented and given finished text; Turkish clips from YODAS, with their captions corrected.
- NaturalVoices, by Zongyang Du, Shreeram Suresh Chandra, Ismail Rasim Ulgen, Aurosweta Mahapatra, Ali N. Salman, Carlos Busso and Berrak Sisman (2025), from podcast episodes by their creators. Each episode has its own licence (list); only CC BY and public-domain episodes were used. Full and light models. One speaker’s turns, segmented, transcribed and given finished text.
- Localized Narratives, by Google (Pont-Tuset et al., 2020). CC BY 4.0. Full model. Voice recordings of people describing images, transcribed and given finished text. The images were not used.
- Common Voice Spontaneous Speech 5.0 (English), by the volunteers of Mozilla Common Voice, through the Mozilla Data Collective. CC0: no credit is needed, and I give it anyway. Full and light models. Spoken answers recorded on people’s own devices; long answers cut into pieces, transcribed and given finished text.
- FLEURS (Turkish), by Google (Conneau et al., 2022). CC BY 4.0. Full and light models. Turkish read speech with its own transcripts, used as they are, so Typd keeps understanding Turkish.
- AMI Meeting Corpus, by the AMI Consortium (I used the University of Edinburgh’s copy). CC BY 4.0. Light model. Headset recordings from the training meetings, with their transcripts turned into finished text.
- VoxPopuli (English), by Meta AI (Wang et al., 2021), from recordings of the European Parliament. CC0, with the Parliament’s legal notice for the recordings. Light model. Parliament speeches with their official transcripts, in an early training round.
- DisfluencySpeech, by Kyra Wang and Dorien Herremans, AMAAI Lab (2024). Apache 2.0, as published. Light model. One speaker acting out real conversations, with fillers and false starts; its transcripts turned into finished text.
Voices for practice dictations
Some training dictations are scripted practice (lists, spoken commands, corrections, fillers), spoken by text-to-speech.
- LibriTTS-R, by Yuma Koizumi and colleagues at Google (2023), restored from LibriTTS and the LibriVox volunteers’ audiobook recordings. CC BY 4.0. Full and light models. One short recording from each of 20 speakers, used only as the voice for synthetic practice dictations.
- Text-to-speech: Kokoro-82M by hexgrad (Apache 2.0), Chatterbox-Turbo by Resemble AI (MIT) and MOSS-TTS-Nano by OpenMOSS (Apache 2.0). Full and light models. They spoke the practice dictations. Their licences don’t ask for credit; I name them anyway.
Transcripts from another model
- Parakeet-TDT-0.6B-v3, by NVIDIA. CC BY 4.0. Full model. Its word-for-word transcripts of the YouTube, podcast and image-description speech above were the starting point for the finished text.
Software inside Typd
- llama.cpp and ggml (MIT), which runs the model. Its Windows build includes LLVM’s OpenMP runtime (Apache 2.0 with LLVM exception).
- Electron (MIT), the app’s framework, with Chromium and its parts under their own licences (listed in LICENSES.chromium.html in Typd’s folder).
- miniaudio (public domain or MIT No Attribution), for the microphone, from Typd 0.2.12.
- uiohook-napi (MIT) and libuiohook (LGPL 3.0), for the dictation key.
- NVIDIA CUDA runtime and cuBLAS (NVIDIA’s CUDA licence), only in the CUDA pack that Typd downloads on a PC with an NVIDIA graphics card.
Typd’s own code is open source under the GPL-3.0: github.com/SameRainbows/typd-app.
Something missing?
If your work is in Typd and missing here, or credited wrongly, write to contact@mehmetdedeler.com and I will fix it.
Last updated: October 2026.