@@ -55,7 +55,9 @@ Two more things make it fast on a phone. Output mostly copies the transcript, so
55
56
## Training
57
58
-Full fine-tune of Qwen3 1.7B, with the embedding and output layers frozen, on about 1,400 handwritten dictation pairs. Each input was also spoken by Kokoro TTS and transcribed by Parakeet TDT 0.6B v2, so the model sees real recognizer output: numbers as digits, stray capitals where audio chunks meet, and misheard words.
59
60
## Limits
61
55
56
## Training
57
58
+Full fine-tune of Qwen3 1.7B, with the embedding and output layers frozen, on about 1,600 handwritten dictation pairs. Each input was also spoken by Kokoro TTS and transcribed by Parakeet TDT 0.6B v2, so the model sees real recognizer output: numbers as digits, stray capitals where audio chunks meet, and misheard words. About 2,000 more examples chain 2 to 4 pairs into one long dictation, because real dictations stack several fixes at once.
59
+
60
+In another-stt, simple rules run before the model: they drop fillers, apply "scratch that", turn "number one," into "1.", and apply corrections that repeat earlier words. Training inputs went through the same rules. Rows whose target the word constraint couldn't produce were left out.
