machine-learning on Purinda Gunasekara https://purinda.com/tags/machine-learning/ Recent content in machine-learning on Purinda Gunasekara Hugo en-AU Sun, 27 Sep 2026 00:10:00 +1000 Fine-tuning Qwen3-ASR for Sinhala – Part 7: The Results https://purinda.com/posts/sinhala-asr-7-results/ Sun, 27 Sep 2026 00:10:00 +1000 https://purinda.com/posts/sinhala-asr-7-results/ The finished model gets 7.1% of characters wrong on Sinhala sentences it never trained on, keeps English within 0.3 points of the original, and runs 17 times faster than real time on a laptop. Also: most of my test sentences turned out to be in the training data. Fine-tuning Qwen3-ASR for Sinhala – Part 6: What's Next https://purinda.com/posts/sinhala-asr-6-whats-next/ Sun, 27 Sep 2026 00:05:00 +1000 https://purinda.com/posts/sinhala-asr-6-whats-next/ What has to be true before the Sinhala model ships, what’s left to do in the app, the questions I can’t answer yet, and the short version of the series. Fine-tuning Qwen3-ASR for Sinhala – Part 5: The First Training Run https://purinda.com/posts/sinhala-asr-5-first-training-run/ Sat, 26 Sep 2026 10:19:00 +1000 https://purinda.com/posts/sinhala-asr-5-first-training-run/ A loss that looked perfect while half the transcripts were wrong, picking a learning rate by watching English get worse, three precision bugs no loss curve would show, and winning English back by blending with the original model. Fine-tuning Qwen3-ASR for Sinhala – Part 4: Training on a Mac, the Same Way the App Runs https://purinda.com/posts/sinhala-asr-4-training-on-a-mac/ Sat, 26 Sep 2026 10:18:00 +1000 https://purinda.com/posts/sinhala-asr-4-training-on-a-mac/ Why the training runs on a MacBook Pro instead of a rented GPU, and the three places where the trainer had to copy the app exactly. One of them was a bug behind 40% of the errors in noisy English. Fine-tuning Qwen3-ASR for Sinhala – Part 3: Preparing the Training Data https://purinda.com/posts/sinhala-asr-3-data/ Sat, 26 Sep 2026 10:17:00 +1000 https://purinda.com/posts/sinhala-asr-3-data/ Rewriting English words into English letters, keeping test speakers out of training, and a replay set labelled by the model itself so it doesn’t forget the languages it already knows. Fine-tuning Qwen3-ASR for Sinhala – Part 2: Why Qwen3-ASR as the Base Model https://purinda.com/posts/sinhala-asr-2-choosing-qwen3-asr/ Sat, 26 Sep 2026 10:16:00 +1000 https://purinda.com/posts/sinhala-asr-2-choosing-qwen3-asr/ Choosing the model to fine-tune: why not one that already knows Sinhala, one model for every language, a decoder that can be told the language, which size, and what starting without Sinhala costs. Fine-tuning Qwen3-ASR for Sinhala – Part 1: Measuring the Models That Already Exist https://purinda.com/posts/sinhala-asr-1-measuring/ Sat, 26 Sep 2026 10:15:00 +1000 https://purinda.com/posts/sinhala-asr-1-measuring/ Before training anything, I measured the speech models that claim to support Sinhala on 300 real recordings. None was usable for dictation, and the reasons shaped everything after.