Daniel Rosehill Blog
Exploring technology, AI, business insights, and personal reflections from a curious mind
3 Kislev 5787
65 hours of my own voice: an index of my speech datasets
64.7 hours and 46,227 takes of my own voice as of 13 November 2026: how the audio datasets on Hugging Face fit together, the hours in each, and the models they feed (Sabich, Fritter, Falafel).
3 Kislev 5787 Update
Hebrish weekend update: why I train the models, how the data is made, and what is next
A video update on Hebrish, my English-Hebrew dictation project: why I fine-tune my own speech models, how the training data is made, the model lineup, and what the project needs next.
3 Kislev 5787
How I record training data for my own speech-to-text models
About 65 hours of my own voice so far: the recording app Claude Code built, a Logitech gamepad as the controller, how the sentences are made, and the speech models it all feeds. With the video.
3 Kislev 5787 Update
Fritter has a new home: English dictation for technical vocabulary
Fritter, my English dictation model for technical vocabulary, moves to its own Fritter-ASR organisation, reaches v1.5 and ships in Anti-Vocale. Plus the version manifest other apps can use.
29 Cheshvan 5787
Meorav: speech synthesis for English-Hebrew code-switching
Meorav is my text-to-speech model for English with Hebrew words dropped in, the way English speakers in Israel talk. Irish accent included. v2.2 is a Qwen3-TTS fine-tune on 5.5 hours of my voice. Before and after samples.
Popular tags
All tags →
28 Cheshvan 5787 Update
Fritter v1.3, a vocabulary cloud and a public tracker
Fritter, my English dictation fine-tune for tech vocabulary, is at v1.3. The demo now opens with a moving cloud of the real terms it was trained on, and the project has a public tracker repo.
27 Cheshvan 5787
One search for skills and MCP tools: introducing Canopy
Canopy is self-hosted progressive discovery for AI agents: one search across skills, plugins and MCP tools, so nothing loads until it's needed. How it works, how it differs, and how to install it.
27 Cheshvan 5787
Forty-four hours of my own voice: my audio datasets on Hugging Face
My speech datasets on Hugging Face, now in one collection and one index: about 44 hours of my own voice, across 39 datasets on my personal account and the Hebrish organisation, and how the hours split.
27 Cheshvan 5787
Hebrish update: 31 hours of audio, and tech vocabulary isn't hurting the Hebrew
Sabich, my English-Hebrew dictation model, has gone from ten hours of audio to 31 and is now a rolling release. Here's the five-part blend, and the first results on whether tech vocabulary costs accuracy on Hebrew words.
26 Cheshvan 5787 Update
Fritter gets a streaming version and direct downloads
Fritter, my English dictation fine-tune for technical vocabulary, is at v1.2 with a first streaming version. Both now download directly from plain URLs, and here is how it differs from my code-switching model.