Keysat
14 lines
1.7 KiB
TypeScript
14 lines
1.7 KiB
TypeScript
import { VersionInfo } from '@start9labs/start-sdk'
|
|
|
|
export const v_0_2_48 = VersionInfo.of({
|
|
version: '0.2.48:0',
|
|
releaseNotes: {
|
|
en_US:
|
|
'Three transcript-quality fixes after the v0.2.45 sort fix didn\'t catch everything. (1) sortAndDedupeTranscript() now also runs on the SINGLE-SHOT transcribe path (previously only on the chunked path) — Gemini occasionally emits entries out of order even within a single API call, especially on flash variants treating long audio as multiple mental "thoughts." Same fix applied to the Hardware (Parakeet) backend single-shot path. (2) New mergeShortEntries() helper coalesces too-granular entries — when the model emits one entry per breath (`► 4:05 um,` / `► 4:07 that is` / `► 4:09 usually based`) producing unreadable transcripts. Merges adjacent entries that are <60 chars AND within 10s of the previous entry\'s timestamp, until each entry contains a complete thought (≥60 chars OR sits at a real speech-pause). Conservative thresholds preserve real short interjections like "Yeah." or "Right." when they\'re meaningful exchanges. (3) Transcribe prompt rewritten to ask for ~30-60s entries containing complete thoughts (2-5 sentences, ~50-200 words) rather than the previous "every 15-30 seconds or at natural pauses" which the model interpreted as "as granular as possible." Also added an explicit "TIMESTAMPS MUST INCREASE MONOTONICALLY" rule plus the [H:MM:SS] format guidance for audio over 60 min. Both gemini.js and hardware.js apply sort + merge to every output path now. Helpers exported from gemini.js, imported into hardware.js to avoid duplication.',
|
|
},
|
|
migrations: {
|
|
up: async ({ effects }) => {},
|
|
down: async ({ effects }) => {},
|
|
},
|
|
})
|