mirror of
https://github.com/stablyai/orca.git
synced 2026-11-16 10:58:10 +00:00
fix(dictation): take the filler language from what the engine is told, and drop the comma before a final filler
This commit is contained in:
@@ -9,6 +9,7 @@ describe('removeFillerWords', () => {
|
||||
['en', 'Yes. Uh. No.', 'Yes. No.'],
|
||||
['en', 'We should ship it, hmm', 'We should ship it'],
|
||||
['en', 'we should um.', 'we should.'],
|
||||
['en', 'We should ship, um.', 'We should ship.'],
|
||||
['en', 'um so we ship', 'so we ship'],
|
||||
['en', 'Um, iPhone builds first.', 'iPhone builds first.'],
|
||||
[
|
||||
@@ -30,6 +31,7 @@ describe('removeFillerWords', () => {
|
||||
['en', 'The umbrella is humming, uh-huh.'],
|
||||
['en', 'Hmm? What did you say?'],
|
||||
['en', 'Ask UM and HMM about it.'],
|
||||
['en', 'Open um.txt and set var_um to 3.'],
|
||||
['ja', 'ええ、そうです。'],
|
||||
['ko', '음이 너무 높아요.'],
|
||||
['zh', '金额不对,那个文件也要改。']
|
||||
@@ -50,14 +52,22 @@ describe('createDictationTextPipeline', () => {
|
||||
expect(clean?.('Äh, um, ship it.')).toBe('Äh, ship it.')
|
||||
})
|
||||
|
||||
it('reads the language from the setting for a multilingual model', () => {
|
||||
it('reads the language from the setting for a multilingual model that takes it', () => {
|
||||
const clean = createDictationTextPipeline(
|
||||
{ language: 'multilingual' },
|
||||
{ language: 'multilingual', dictationLanguages: ['de', 'en'] },
|
||||
{ dictationLanguage: 'de' }
|
||||
)
|
||||
expect(clean?.('Äh, um fünf Uhr.')).toBe('Um fünf Uhr.')
|
||||
})
|
||||
|
||||
it('ignores a language setting the model ignores', () => {
|
||||
const clean = createDictationTextPipeline(
|
||||
{ language: 'multilingual' },
|
||||
{ dictationLanguage: 'en' }
|
||||
)
|
||||
expect(clean?.('Er hat um zehn Uhr angerufen.')).toBe('Er hat um zehn Uhr angerufen.')
|
||||
})
|
||||
|
||||
it('leaves text untouched when filler removal is off', () => {
|
||||
expect(createDictationTextPipeline({ language: 'en' }, { removeFillerWords: false })).toBeNull()
|
||||
})
|
||||
|
||||
@@ -17,13 +17,16 @@ FILLERS_BY_LANGUAGE.yue = FILLERS_BY_LANGUAGE.zh
|
||||
// For an unknown language: English fillers that are no word elsewhere ("um" is German and Portuguese, "er" German).
|
||||
const LANGUAGE_NEUTRAL_FILLERS = 'u+h+m*|u+m{2,}|e+r+m+|h+m+|m{3,}'
|
||||
|
||||
const WORD_CHAR = String.raw`[\p{L}\p{N}\p{M}'’\-]`
|
||||
// A filler stands alone: no word character touches it, and "hm?" or "eh!" is a question or a call, so it stays.
|
||||
const WORD_CHAR = String.raw`[\p{L}\p{N}\p{M}_/'’\-]`
|
||||
// A filler stands alone: nothing joins it to a word (a dot only does inside a name like "um.txt"),
|
||||
// and "hm?" or "eh!" is a question or a call, so it stays.
|
||||
const STANDALONE = (fillers: string): string =>
|
||||
String.raw`(?<!${WORD_CHAR})(${fillers})(?!${WORD_CHAR})(?!\s*[??!!])`
|
||||
String.raw`(?<!${WORD_CHAR}|${WORD_CHAR}\.)(${fillers})(?!${WORD_CHAR}|\.${WORD_CHAR})(?!\s*[??!!])`
|
||||
// The comma or ellipsis after a filler goes with it.
|
||||
const TRAILING_PAUSE = String.raw`(?:\s*(?:[,,、…]|\.{3}))?\s*`
|
||||
const SENTENCE_END_RE = /^[.!?。!?]+\s*/u
|
||||
// A comma left in front of a removed filler that ended the sentence or the text.
|
||||
const TRAILING_SOFT_PUNCTUATION_RE = /[\s,,、]+$/u
|
||||
const SENTENCE_START_RE = /(^|[.!?。!?])\s*$/u
|
||||
const NON_LATIN_LETTER_RE = /(?=\p{L})\P{Script=Latin}/u
|
||||
|
||||
@@ -75,11 +78,11 @@ export function removeFillerWords(text: string, languages: readonly string[] | u
|
||||
cursor = match.index + match[0].length
|
||||
const atSentenceStart = SENTENCE_START_RE.test(out)
|
||||
const sentenceEnd = SENTENCE_END_RE.exec(text.slice(cursor))
|
||||
// "we should um." keeps its full stop; "Um." on its own goes entirely.
|
||||
// "we should, um." keeps its full stop; "Um." on its own goes entirely.
|
||||
if (sentenceEnd && atSentenceStart) {
|
||||
cursor += sentenceEnd[0].length
|
||||
} else if (sentenceEnd) {
|
||||
out = out.trimEnd()
|
||||
out = out.replace(TRAILING_SOFT_PUNCTUATION_RE, '')
|
||||
}
|
||||
// Only a capitalised "Um" began a sentence the next word now begins.
|
||||
capitalizeNext ||= atSentenceStart && filler.charAt(0) !== filler.charAt(0).toLowerCase()
|
||||
@@ -90,5 +93,5 @@ export function removeFillerWords(text: string, languages: readonly string[] | u
|
||||
// A filler at the very end leaves the comma before it dangling.
|
||||
const endsOnFiller = cursor === text.length
|
||||
keep(text.slice(cursor))
|
||||
return (endsOnFiller ? out.replace(/[\s,,、]+$/u, '') : out).trim()
|
||||
return (endsOnFiller ? out.replace(TRAILING_SOFT_PUNCTUATION_RE, '') : out).trim()
|
||||
}
|
||||
|
||||
@@ -1,24 +1,31 @@
|
||||
import { DICTATION_LANGUAGE_AUTO } from './dictation-language'
|
||||
import { engineDictationLanguage } from './dictation-language'
|
||||
import { removeFillerWords } from './dictation-filler-words'
|
||||
import type { SpeechModelManifest, VoiceSettings } from './speech-types'
|
||||
|
||||
export type DictationTextSettings = Pick<VoiceSettings, 'dictationLanguage' | 'removeFillerWords'>
|
||||
|
||||
/** The languages the dictated text is in: a single-language model's own, else the user's setting; undefined when unknown. */
|
||||
export function dictatedTextLanguages(
|
||||
model: Pick<SpeechModelManifest, 'language'> | undefined,
|
||||
type DictationTextModel = Pick<SpeechModelManifest, 'language' | 'dictationLanguages'>
|
||||
|
||||
/** The languages the dictated text is in: a single-language model's own, else the one the engine is told; undefined when unknown. */
|
||||
function dictatedTextLanguages(
|
||||
model: DictationTextModel | undefined,
|
||||
setting: string | undefined
|
||||
): string[] | undefined {
|
||||
if (model && model.language !== 'multilingual') {
|
||||
if (!model) {
|
||||
return undefined
|
||||
}
|
||||
if (model.language !== 'multilingual') {
|
||||
// Bilingual models name both, as in 'zh-en'.
|
||||
return model.language.split('-')
|
||||
}
|
||||
return setting && setting !== DICTATION_LANGUAGE_AUTO ? [setting] : undefined
|
||||
// The Voice pane disables a setting the model ignores, so a stale one says nothing about what was spoken.
|
||||
const language = engineDictationLanguage(model, setting)
|
||||
return language ? [language] : undefined
|
||||
}
|
||||
|
||||
/** The passes dictated text goes through before anyone sees it, or null when every pass is off. */
|
||||
export function createDictationTextPipeline(
|
||||
model: Pick<SpeechModelManifest, 'language'> | undefined,
|
||||
model: DictationTextModel | undefined,
|
||||
settings: Partial<DictationTextSettings> | undefined
|
||||
): ((text: string) => string) | null {
|
||||
if (settings?.removeFillerWords === false) {
|
||||
|
||||
Reference in New Issue
Block a user