Tag
fine-tuning
14 posts · all tags
24 Cheshvan 5787
Improving a personal ASR model by using it every day
How to improve a personal speech-to-text model by using it daily: an agent reviews your transcript history, you confirm the errors, and small fine-tunes follow. A playbook from Parakeet and Hebrish, with a PDF and a Claude Code skill.
23 Cheshvan 5787
Datasets for fine-tuning speech recognition on your own voice
A guide to my public Hugging Face datasets of read-aloud sentences for fine-tuning speech recognition: tech, Linux, smart home, AI agents, homophones, Hebrew code-switching and voice notes. What each holds and which to use.
23 Cheshvan 5787
Contextual Sentence Creator: a 3B model that writes ASR training sentences
A Granite 4.2 3B fine-tune that writes natural sentences using a given term, for recording speech recognition training data. 95.7% of sentences pass the acceptance check; GGUF build included.
23 Cheshvan 5787
Fritter: a Parakeet fine-tune for English tech dictation
Fritter is my Parakeet TDT 0.6B v2 fine-tune for English dictation without Hebrew: tech terms, numbers and brand names. Tech WER falls from 7.53% to 4.01% on my voice. Browser demo and GGUF builds.
23 Cheshvan 5787
S1-ASR-Cleanup: cleaning up dictation with a fine-tuned S1-mini
A fine-tune of S1-mini by Superwhisper that cleans speech-to-text dictation: paragraphs, dates and money as digits, contractions and fillers handled. 0.6B, with GGUF builds for llama.cpp and phones.
10 Cheshvan 5787
Adding paragraph breaks to speech-to-text output with Gemma 3 270M
A Gemma 3 270M fine-tune that adds paragraph breaks to flat speech-to-text output and changes nothing else. Small enough for on-device use, with an ONNX export. Break F1 0.78 on 26 held-out memos.
10 Cheshvan 5787
Bolding the key lines of a dictated email with Gemma 3 270M
A preliminary Gemma 3 270M fine-tune that bolds the key lines of a dictated email (the asks, the purpose, the key facts) and changes nothing else. Small enough for on-device use. Word F1 0.62 on six held-out emails.
10 Cheshvan 5787
Jachnun and Sabich: the first stable Hebrish models, in depth
A deep dive into the first stable Hebrish models, Jachnun v1.0 and Sabich v1.0: the data, the training recipe, the results against stock Parakeet and Whisper, the formats, and the full white paper.
10 Cheshvan 5787
Hebrish: one home for my English-Hebrew speech models
Hebrish is my one-person, not-for-profit home for English-Hebrew speech models: Jachnun, Sabich, Shakshuka and the Whisper fine-tunes, free apps, open data and benchmarks, all at hebrish.app.
10 Cheshvan 5787
Proximate Pseudonymizer: a small model that swaps identifying details for close substitutes
A proof-of-concept Qwen3-0.6B fine-tune that rewrites text with names, places and dates swapped for close, consistent substitutes (Daniel to David, Jerusalem to Tel Aviv). Runs on CPU. Still leaks 9.3% of details, so review every output.
9 Cheshvan 5787
Boxty v1.0: English dictation without the Hebrew
Boxty is the no-Hebrew sibling of my Jachnun code-switching model: Parakeet fine-tuned on my voice for plain English and tech dictation. Tech WER drops from 2.51% (stock) to 1.35%, and general English holds.
8 Cheshvan 5787
Voice typing that understands Hebrish
Speech-to-text mangles the Hebrew words English speakers in Israel use every day. Hebrish fixes that: free on-device apps for Android, Linux and Windows, the Jachnun and Sabich models, results, and how you can help.
6 Cheshvan 5787 Update
Whisper Hebrish v2 Released
Version 2 of Whisper Hebrish, my Whisper fine-tune for English with Hebrew words mixed in. Code-switching WER drops from 14.7% (stock turbo) to 6.7%, and there are whisper.cpp builds.
7 Nisan 5786
Fine-tuning Whisper on Modal's serverless GPUs
A script for fine-tuning OpenAI's Whisper speech recognition models using Modal's serverless GPU infrastructure.