feat: vocabulario de dominio, reconhecimento on-device e captura de debug#2
Open
jcelos wants to merge 1 commit intoDeveloperParana:mainfrom
Open
feat: vocabulario de dominio, reconhecimento on-device e captura de debug#2jcelos wants to merge 1 commit intoDeveloperParana:mainfrom
jcelos wants to merge 1 commit intoDeveloperParana:mainfrom
Conversation
…ebug
Tres melhorias entrelacadas pra qualidade da transcricao:
* Vocabulario por sala + correcao fuzzy: dicionario de termos tecnicos
(IA, LLM, Claude, Next.js, etc.) editavel via modal por sala. Engine
pede N alternativas ao Web Speech, ranqueia por hits no vocabulario,
e aplica correcao fuzzy (Levenshtein) pos-reconhecimento. Filtro de
ruido descarta finais curtos de baixa confidence (lista de fillers).
* Reconhecimento on-device (Chrome 139+): detecta SR.available e
dispara SR.install dentro do callstack do click do usuario pra
preservar transient activation. Sem isso o Chrome lanca SecurityError
("Requires handling a user gesture when availability is downloadable")
e o install nunca acontece. Quando install resolve, troca cloud->local
sem perder audio. Validado contra Chrome 147 headless via probe CDP
com 4 cenarios (downloadable+OK, available, install-rejected, Chrome
antigo). UI mostra badge "On-device"/"Cloud" e botao manual de retry
quando o gesture expira.
* Captura de debug: botao na barra de controle grava audio (.webm) +
log estruturado de eventos do engine (alternativas, picked, corrigido,
emitido, processLocally) em JSON. scripts/analyze-debug.mts cospe
metricas de latencia interim->final, hits do vocab, distribuicao de
confidence, drops por filtro de ruido, duplicacao entre sessoes.
Usado pra diagnosticar bugs de vocabulario e medir latencia real do
cloud STT (~5s p50) vs alvo on-device (~500ms).
Speech engine modularizado em lib/speech/{types,web-speech,index,debug-capture}
pra permitir engines alternativos (Whisper, Deepgram, Vosk) sem mexer
no chamador. Cursor rule documenta caveats descobertos do API on-device
e como rodar os probes E2E pra regredir mudancas no fluxo.
|
@jcelos is attempting to deploy a commit to the sschonss' projects Team on Vercel. A member of the Team first needs to authorize it. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Resumo
Três melhorias entrelaçadas pra qualidade da transcrição:
SR.availablee disparaSR.installdentro do callstack do click do usuário pra preservar transient activation — sem isso o Chrome lançaSecurityError("Requires handling a user gesture when availability is downloadable") e o install nunca acontece. Quando install resolve, troca cloud→local sem perder áudio. UI mostra badgeOn-device/Cloude botão manual de retry quando o gesture expira..webm) + log estruturado de eventos do engine (alternativas, picked, corrigido, emitido,processLocally) em JSON.scripts/analyze-debug.mtscospe métricas de latência interim→final, hits do vocab, distribuição de confidence, drops por filtro de ruído, duplicação entre sessões.Speech engine modularizado em
lib/speech/{types,web-speech,index,debug-capture}pra permitir engines alternativos (Whisper, Deepgram, Vosk) no futuro sem mexer no chamador. Cursor rule (.cursor/rules/realtime-transcription.mdc) documenta caveats descobertos do API on-device e como rodar os probes E2E pra regredir mudanças no fluxo.Validação
tsc --noEmitlimpo,npm run buildok (176 kB no/sala/[code]/mestre).downloadable+ install OK → badgeOn-deviceavailableimediato → badgeOn-devicesem chamar installBaixar modelo on-deviceSR.available→ sem badge, sem crashTest plan
Baixando modelo...→On-device.Claude,LLM,Next.js) e conferir correção.localStorage..webm+.json, rodarnpx tsx scripts/analyze-debug.mts arquivo.json.Cloud; transcrição segue funcionando sem regressão.