@lacspace/speakable

Say it right: the spoken-form layer every Nepali/English TTS pipeline is missing. Rewrites what engines mangle — numbers, Bikram Sambat dates, times, currency, percentages, phone numbers, plates, units, ordinals, acronyms, abbreviations, cross-script names — before any engine (edge-tts, Piper, Azure, Google…) sees the text. Then segments it with pauses and rate, emits SSML, estimates word timings, and aligns the engine's own word boundaries back to the original text so captions never drift.

npm i @lacspace/speakable

import { speakable, spokenText, alignWordBoundaries, originalWordTimings } from "@lacspace/speakable";

spokenText("२०८३ असोज १६ गते रु. १ लाख ५० हजार, १५.२% वृद्धि, समय १३:३८, NEPSE २,६५०.१२", { lang: "ne" });
// → "दुई हजार त्रियासी साल असोज सोह्र गते एक लाख पचास हजार रुपैयाँ, पन्ध्र दशमलव दुई प्रतिशत वृद्धि,
//    समय दिउँसो एक बजेर अठतिस मिनेट, नेप्से दुई हजार छ सय पचास दशमलव एक दुई"

spokenText("NEPSE rose 12.5 points (0.47%) on 1 October 2026; call +977-9841234567 by 6:30 PM.", { lang: "en" });
// → "nepsay rose twelve point five points (zero point four seven percent) on the first of October,
//    twenty twenty-six; call plus nine seven seven, nine eight four one, two three, four five, six seven by six thirty p m."

const r = speakable(article, { lang: "ne", voice: "ne-NP-HemkalaNeural", pronunciations: { Lamichhane: { ne: "लामिछाने" } } });
r.ssml                 // <speak><s><prosody rate="-8%">…</prosody></s><break time="400ms"/>…</speak>
r.text                 // plain spoken text for engines without SSML
r.segments             // sentences: { text, orig, pauseAfter, rate, durationMs, words:[{word,startMs,endMs}] }
r.estimatedDurationMs  // before you even call the engine

// After synthesis, feed edge-tts WordBoundary events ({ text, offsetMs, durationMs }) back:
const aligned = alignWordBoundaries(r.tokens, r.segments, events);
originalWordTimings(aligned);  // [{ orig: "रु. १ लाख", startMs: 0, endMs: 880, … }] — caption units in the ORIGINAL text

What it normalises

| input | ne | en | |---|---|---| | २०८३ असोज १६ गते / 16 Asoj 2083 / 2083-06-16 | दुई हजार त्रियासी साल असोज सोह्र गते | sixteenth of Asoj, two thousand and eighty-three | | 1 October 2026 / 2026-10-02 | एक अक्टोबर दुई हजार छब्बिस | the first of October, twenty twenty-six | | १३:३८ / 6:30 PM | दिउँसो एक बजेर अठतिस मिनेट / साँझ छ बजेर तीस मिनेट | one thirty-eight / six thirty p m | | रु. १ लाख ५० हजार / NPR 1,20,000 / Rs 197.5 | एक लाख पचास हजार रुपैयाँ / … / एक सय सन्तानब्बे रुपैयाँ पचास पैसा | one lakh fifty thousand rupees … | | $120 million | बाह्र करोड डलर | one hundred and twenty million dollars | | १.५ लाख / २.५ करोड / 1.3 लाख | डेढ लाख / अढाई करोड / एक लाख तीस हजार | one lakh fifty thousand … | | १५.२% | पन्ध्र दशमलव दुई प्रतिशत | fifteen point two percent | | ९८४१२३४५६७ / 01-4412345 | नौ आठ चार एक, दुई तीन, … (digit by digit) | nine eight four one, … | | बा १२ प ३४५६ | बा बाह्र प तीन चार पाँच छ | Ba twelve Pa three four five six | | १६औँ / 16th | सोह्रौँ | sixteenth | | 25°C, 5 km, 200 MW | पच्चिस डिग्री सेल्सियस, पाँच किलोमिटर, दुई सय मेगावाट | twenty-five degrees Celsius … | | NEPSE, NRB, KMC, एनएस, UNESCO | नेप्से, एन आर बी, के एम सी, एन एस, युनेस्को | nepsay, N R B, K M C, … | | डा. वि.सं. Dr. Rs. | डाक्टर, विक्रम संवत्, … | Doctor, Bikram Sambat, … | | Sandeep Lamichhane in Nepali text | सन्दीप लामिछाने (translit) or your override | — | | काठमाडौं in English text | — | Kathmandau (translit) or your override |

Nepali has a unique word for every number 0–99 (NE_0_99), Indian scales (हजार/लाख/करोड/अर्ब/खर्ब) and fraction idioms (आधा, डेढ, अढाई, सवा, साढे, पौने) — all built in. English picks western or Indian grouping automatically from context (लाख/crore/रु/Rs present → lakh/crore), or set grouping.

Pacing, SSML, timing

• Sentences split on ।/./!/? (abbreviations are expanded first, so डा. never ends a sentence); blank lines are paragraph pauses.
• Each sentence gets pauseAfter (400 ms sentence, 700 ms paragraph, 180 ms at commas/clauses) and a rate that drops toward 0.88 as the share of numeric words rises — numbers are read slower, as a newsreader does.
• ssml uses <s>, <break>, <prosody rate> (the dialect edge-tts and Azure accept); text is the plain fallback.
• Word timings are estimated from syllable counts (syllables()), tuned for Nepali (≈125 wpm) and English (≈160 wpm) neural voices; estimatedDurationMs lets you budget a video before synthesis.
• alignWordBoundaries(tokens, segments, events) snaps the engine's real WordBoundary events onto the spoken words (robust to split/merged words) and maps them to original spans; originalWordTimings() collapses to one entry per original token for karaoke/burn-in captions.

Options

lang ("ne" | "en" | "auto"), pronunciations ({ word: { ne, en } | string }), acronyms (extend the table), foreignWords ("transliterate" default | "keep"), dateSystem ("auto" → Nepali month or ISO year ≥ 2050 = BS), dateStyle ("british" | "american"), grouping, voice, baseRate, wpm, sentencePauseMs, clausePauseMs, paragraphPauseMs, numberRate.

describe() returns the command schema for an AI conductor. Zero third-party deps (uses @lacspace/translit for cross-script names).

Licence

Lacspace Free Licence v1.0 — free for personal and commercial use.