How accurately does speech recognition place synchronization units inside a larger narration chunk?

Aberta
#325 0 comentários 0 reações 0 responsáveis Ver no GitHub

Ninguém assumiu esta issue ainda.

Avaliação

Dificuldade
5/5
Tempo estimado
Mais de uma semana
Facilidade para iniciantes
35/100
Tipo de issue
Funcionalidade
Clareza
Claramente especificada
Status de atividade
Ativa
Stack de tecnologia
typescript
Domínio
audio-video-rtc

Direção de pesquisa

Start with the prototype's word alignment of recognized words to the narration document's words, then review issues #316, #322, #323, and #324 for prior measurements and the chunk size. Test Whisper large v3 turbo and, if available, the other candidate across the specified languages and fixture articles; measure placement accuracy, chunk splitting effects, cost, and latency, settling disagreements by ear. Report measured versus unverified claims and propose an acceptance rule, including whether one method covers every language.

Escrita pelo modelo de indexação a partir do texto da issue.

Descrição

wayfinder:research

Question

How accurately does speech recognition with word timing place each synchronization unit's start inside a larger narration chunk's audio?

Overlap and cut needed one cut point per audio segment, and Whisper large v3 turbo placed 116 of 116 researched cuts (Which aligner finds overlap-and-cut points reliably?). In production that did not hold, and overlap and cut was withdrawn on 2026-11-13 (ADR 0028): Whisper often started a chunk's first word early, stretching it back over the pause in front of it, so the deterministic search never saw the real pause. In 11 live re-runs of one production boundary, 2 cut checks failed and 3 accepted a cut inside the previous sentence. A unit start is exactly that spot, so this question must not take Whisper's word starts as the unit starts. Decoupling needs one placement per synchronization unit inside a chunk of the size How long can one Gemini 3.8 TTS request get before a fresh start inside it? settles, so errors add up per unit, and recognition drift over several minutes of audio may grow. Using that chunk size, the fixture articles in English, German and French, compare two placements: the prototype's deterministic word alignment of recognized words to the narration document's words, and a model-based placement, an AI step that maps the recognized words, their timing and the audio's pauses onto the synchronization units, which is the operator's preferred direction (2026-11-13). For each, measure:

  • how often a unit start lands in the real pause before its first word, as the audio's silence shows it rather than Whisper's word start, versus inside a word or in the wrong pause, including at headings, the source title and source subtitle, list items, table cells, numbers, abbreviations and names;
  • whether long chunks need the audio split before recognition, and what that does to word timing;
  • speech recognition cost and latency per chunk against synthesis.

Beyond English, German and French, take a harder sample (headings, numbers, table cells, short units) in several of the 28 narration languages Whisper does not list, and compare two placements there: Whisper matched to the narration text letter by letter, which placed 36 of 36 sentence starts on an easy sample (Which speech recognition models with word timing cover Cup's narration languages?), and forced alignment with Omnilingual ASR CTC 300M, which agreed with Whisper at 38 of 39 unit starts and missed only at numbers (How can synchronization units be placed in audio without speech recognition?). Where the two disagree, settle it by ear. The answer should say whether one method can serve every narration language it covers, so the uncovered languages need no second pipeline.

Use Whisper large v3 turbo, plus the best other candidate from Which speech recognition models with word timing cover Cup's narration languages? if it is resolved by then. Propose an acceptance rule for a placement that also catches a placement inside the previous sentence, which overlap and cut's cut check accepted. The 14 recorded Gemini takes with verified pauses at 424d7b2e libs/audiobook-production/test/evals/cases/ are ready-made fixtures for unit starts at headings, bylines, numbers, table cells and quotes, and recognize-speech.ts at that commit is a working Whisper call through the Workers AI binding. Mark measured and unverified claims; light paid probes are fine.

Part of Map: decouple speech synthesis from synchronization units.

Linguagem predominante
TypeScript
Estrelas
1
Forks
0
Métricas de merge de PRs
Nenhum PR com merge em 30d

Preparar o ambiente

Este projeto não oferece contêiner de desenvolvimento, Dockerfile nem guia de contribuição, então a configuração fica por sua conta: comece pelo README e veja nosso guia da primeira contribuição para os passos gerais.

Primeiros passos

  1. Leia a issue inteira e depois o guia de contribuição do projeto.
  2. Comente na issue dizendo que vai assumir — evita que duas pessoas façam o mesmo trabalho.
  3. Faça um fork do repositório e trabalhe em uma branch.
  4. Abra um pull request que referencie o número da issue.

Mais de patricktree/cup

Todas as issues de patricktree/cup

Issues semelhantes

Mais issues de TypeScript

Receba novas issues na sua caixa de entrada

Um resumo curto de issues do GitHub para quem está começando.