Lex

import { Lex } from 'alkove'

const lex = await Lex.create()
await lex.phonemize('Camille', 'fr')
// [{ ipa, p }, …] up to 3 pronunciations, the likeliest first

alkove-lex-1-preview turns a written word into its most likely pronunciations, with their probabilities, in the phoneme set of Spot. Names, brands and invented words included. It is what lets Spot listen for any word you type.

Size	1.8 MB (int8 weights)
Languages	French, English, Spanish, Portuguese, Italian, Catalan, Romanian
Input	a word of up to 33 characters, and its language
Output	up to 3 pronunciations, in Spot’s 58 phonemes, each with its probability
License	Apache-2.0

Using Lex from the SDK

Most apps never call Lex directly: Spot.compile() and Spot.enroll() run it for you when you give them text.

import { Spot } from 'alkove'

const word = await Spot.compile('Hey Nova', { lang: 'en' }) // Lex runs here

Call it yourself to show users how a word will be heard, before they pick it as a wake word, or to check two words do not sound alike:

import { Lex } from 'alkove'

const lex = await Lex.create()
const [best] = await lex.phonemize('Nova', 'en')
const options = await lex.phonemize(['Nova', 'Noah'], 'en') // one list per text

Lex.distance(options[0][0].ipa, options[1][0].ipa) // 0 = same sound, 1 = nothing in common
Lex.langs // ['fr', 'en', 'es', 'pt', 'it', 'ca', 'ro']
Lex.inventory // the 58 phonemes, in IPA

API	What it gives
Lex.create()	a Lex instance, once the model is downloaded and loaded in the background worker
lex.phonemize(text, lang)	the pronunciations of text, { ipa, p }, likeliest first. A phrase works too: each word’s pronunciations are combined, the 3 likeliest kept.
lex.phonemize(texts, lang)	the same for several texts, one list each
Lex.distance(a, b)	how different two pronunciations sound, from 0 (same) to 1; phonemes that are easily confused count half
Lex.langs	the supported languages
Lex.inventory	Spot’s phonemes, in IPA

An unsupported language rejects with an AlkoveError.

Writing the phonemes yourself

If Lex gets a name wrong, write the pronunciation between slashes, in IPA from Lex.inventory. Lex is then not used at all:

const word = await Spot.compile('/kamij/', { lang: 'fr' })

Phonemes outside the inventory are refused.

Text that Lex reads poorly

When it compiles a word, the SDK warns about text that will not be read the way you expect. The warnings are on word.warnings:

Code	Why
text.digits	Numbers are not read: write them in letters.
text.acronym	Acronyms are read as words: write them as they sound.
text.abbreviation	Abbreviations are read as written.
word.too-short	Short words give more false alarms in Spot.

Performance

Native, 1 thread, Galaxy Z Fold8 (Android 17)	about 2.9 ms per word: 1.1 ms to read it, then 0.19 ms per step of the search
Native, 1 thread, Apple M1 Max (macOS)	about 4.1 ms per word: 1.6 ms to read it, then 0.27 ms per step of the search

Lex runs in the background worker, never on the audio path.

Benchmarks

On 300 held-out words across the 7 languages, the first pronunciation is right for 86.0 % of them.

Limitations

• Words of up to 33 characters; the 7 languages above.
• Loanwords and names follow the rules of the chosen language. For a name said in two languages, compile it with both: Spot.compile('Nova', { lang: ['fr', 'en'] }).

Out of scope: phonetizing whole sentences for speech synthesis (it works word by word, up to 33 characters), languages outside the 7 above, and phoneme sets other than Spot’s 58 phonemes.

Running it without the SDK

The model is an .alk file for the Alkove engine (Rust, Apache-2.0), natively or in WebAssembly; its format is documented. It holds two plans over the same weights. The encoder plan reads the word once. Then the decoder plan gives, one call per position, the probabilities of the next phoneme for each of 3 beams, and keeps each beam’s keys and values in its state. You run a beam search around it. Whole numbers are held as floats.

Encoder	Type
word (in)	float32 [1, 33]	the word’s Unicode code points after NFC, lowercase and ’ → ', zeros after
lang (in)	float32 [1]	index of the language in the manifest’s langs
cross (out)	float32 [3, 34, 256]	for each decoder layer, the keys then the values over the word
pad (out)	float32 [34]	1 where there is no letter

Decoder	Type
token (in)	float32 [3]	each beam’s phoneme at pos: bos at 0
parent (in)	float32 [3]	the row of the previous call each beam goes on from: zeros at 0
pos (in)	float32 [1]	the position, from 0
cross, pad (in)		the encoder’s outputs, the same for the whole word
logp (out)	float32 [3, 61]	log-probabilities of the phoneme after pos, for each beam

Clear the decoder’s state before each word. The model card gives the beam search step by step. The SDK runs it, combines the words of a phrase and merges duplicate pronunciations for you.