Tongue

Name the language from three words.

On-device language identification for short text across 84 languages.

Install

Swift (requirements)

.package(url: "https://github.com/Desert-Ant-Labs/desert-ant-core.git", from: "3.1.0")

Then add the Tongue product to your target.

Kotlin (requirements)

implementation("ai.desertant:tongue:3.1.0")

JavaScript (requirements)

npm i @desert-ant-labs/tongue

Usage

Nothing to download and nothing async: the 2 MB model ships inside the package, and a detection is pure arithmetic.

Swift

import Tongue

let tongue = try Tongue()                      // loads the bundled 2 MB model
let detection = tongue.detect("kann ich das haben")

detection.language          // "de"
detection.reliability       // .confident
detection.candidates        // [Prediction(language: "de", probability: 0.999…), …]
detection.isTooCloseToCall  // false

Kotlin

Tongue is a plain jar rather than an AAR, a pure Kotlin port with no native libraries, so it also runs on a bare JVM (17+).

import ai.desertant.tongue.Tongue

// Android: pass the Context. On a bare JVM call Tongue.bundled().
val tongue = Tongue.bundled(context)
val detection = tongue.detect("kann ich das haben")
detection.language                               // "de"
detection.isTooCloseToCall                       // false

JavaScript

One import everywhere: no wasm, no LiteRT.js, no native core.

import { Tongue } from "@desert-ant-labs/tongue";

const tongue = await Tongue.load();                    // Node: reads the bundled model
const detection = tongue.detect("kann ich das haben");
detection.language;                                    // "de"
detection.isTooCloseToCall;                            // false

In a browser, serve tongue's two model files yourself (a bundler does not serve files out of node_modules) and pass from; both files are exported subpaths, so a copy script can require.resolve them under pnpm and Yarn PnP too:

const tongue = await Tongue.load({ from: "/models/tongue" });

Files

There is no tokenizer file. tongue hashes raw character n-grams, so nothing has to be shipped or version-matched alongside the weights.

Architecture

Two stages. No encoder, no learned tokenizer, no per-language models.

• Script router (zero parameters). A UAX #24 script table decides any text whose script belongs to one language outright, Hangul to Korean, Greek to Greek, Thai to Thai, and narrows multi-language scripts (Cyrillic, Arabic, Devanagari, Bengali) to their candidate sets before the model runs. Presence beats dominance, so a Latin brand name embedded in Greek text does not derail the route.
• Lexical model. FNV-1a-hashed character n-grams (orders 1-5, marked with word boundaries) over Unicode scalars, summed through an int8 EmbeddingBag into a linear head, decoded with a per-script masked softmax. The hash table is the feature space, so model size does not grow with the language count.
• Prior correction. The training corpus caps large languages and under-fills thin ones, so the head absorbs a label prior. A fixed -tau*log(prior) shift (tau = 0.75) is folded into the exported bias, which costs nothing at runtime and keeps confusion pairs from collapsing onto the better-resourced side.
• Calibrated abstention. Reliability is keyed off input length and the margin between the top two candidates, not raw softmax confidence, which is overconfident on very short text. Below the threshold the model reports a tentative answer or a tie instead of committing.

Inputs and outputs

Input: a short UTF-8 string. The runtime normalizes it (NFC, lowercase, URLs / mentions / digits stripped, 512-character cap), hashes n-grams, and consults the router before the graph. Output: ranked ISO 639-1/639-3 codes with probabilities, plus a reliability signal (confident / likely / tentative). Script-decided inputs return a single confident answer.

The ONNX graph carries only the head: values (int64 hashed bucket ids) and offsets (int64 per-sample starts) in, logits out. The normalizer, hasher and router are reimplemented natively per platform and verified against golden vectors: they must run before the model is consulted, and on inputs the model never sees. tongue_meta.json documents both tables.

Coverage

59 languages are learned by the lexical model and a further 25 are decided by script alone, across 31 scripts (Latin, Cyrillic, Arabic, Greek and the CJK and Indic families among them) for 84 languages in total. One of the 84, Mongolian, is detected only in the traditional Mongolian script; see failure mode 3.

Sizes

Two quantisations of the same weights ship side by side. Pick on bundle budget, not on principle.

int4 uses one scale per embedding dimension with the scale clipped at the 99th percentile: plain max-scaling spends most of the 16 levels on a handful of outliers and costs 3.8pp instead of 0.4pp. The loss lands almost entirely on one- and two-word input, where there are fewer n-grams for quantisation error to cancel across; full sentences are unaffected within measurement noise. Since short text is what this model is for, int8 stays the default and int4 is the option when a megabyte matters more than the last half point.

Failure modes (read before deploying)

Publishing these is part of the product.

1. One or two words is often genuinely undecidable, and no model size fixes it. A single common word frequently belongs to several languages at once ("sale" is English, French and Italian; "la casa" is equally Italian and Spanish). tongue reports a tie or a tentative answer in these cases. It does not catch every one: a phrase mixing languages, like "un garage sale", can still draw a confident-looking single answer. Mitigation: treat low-reliability output as "unknown", not as an answer, and ask for more text where the product allows it.

2. Malay and Indonesian are not reliably separable. They share vocabulary and orthography to the point where short samples carry no distinguishing signal. This is a structural limit, not a tuning gap, and it is not cheaply closable at this size, and every detector we measured struggles with it. Mitigation: if you need the distinction, treat ms/id as one bucket or disambiguate from user locale.

3. Mongolian is detected only in the traditional Mongolian script. Mongolian written in Cyrillic, the dominant modern orthography, is not distinguished from the other Cyrillic languages and will usually come back as Russian. The language count includes Mongolian because the traditional script works; Cyrillic Mongolian does not. Mitigation: do not rely on tongue for Cyrillic Mongolian.

4. Brand names, numbers and code are not language. "Samsung Galaxy", "v1.2.3" and "2024 annual report" have no correct answer; the model will still return its best guess for anything with letters in it. Mitigation: filter non-prose input before detection.

5. Single-word scores are vocabulary recognition, not generalization. The frequent words of a language appear in everyone's training data, so any detector's single-word accuracy partly measures memorized vocabulary. Read the word-pair and sentence numbers as the generalization signal.

Measured quality (the shipped artifact, not the checkpoint)

Every number below is measured on the shipped int8 weights, on three public benchmarks, with other detectors run on the identical rows and language subsets. Higher is better.

FLORES-200, a benchmark none of these detectors trained on

Sentences from FLORES-200 truncated to their first 2, 3 and 5 words. Accuracy over the 20 languages the three detectors share.

The lingua test set, the benchmark that library publishes

1,000 single words, word pairs and sentences per language, drawn from the same collection lingua trains on. Accuracy over the languages we share.

eld, an independent benchmark, held out from training

Accuracy over the languages tongue supports (53,035 single-word rows, 53,613 word pairs, 53,141 sentences, 9,066 tweets). Apple is the built-in system detector; HeLI-OTS is a 51 MB JVM model. lingua 2.2.0 installs as a single 293 MB compiled extension with its language models embedded.

How these numbers were made

• Benchmarks are eval-only. FLORES-200, WiLI-2018 and the eld benchmark are never trained on. Leipzig/Wortschatz corpora are excluded from training in every form, because a competing detector's published test set is drawn from them.
• Evaluation splits are leakage-controlled. The training corpus is split along the Tatoeba translation-link graph, so a sentence and its translations cannot straddle train and validation.
• Numbers are re-measured on the exported bytes, not extrapolated from the training checkpoint, and the pure-JavaScript runtime is verified against the Python reference on golden vectors (currently 119/119 identical, worst probability delta 1.1e-16).
• Latency is measured per single detection in JavaScript on an Apple-silicon laptop: 0.013 ms for one word, 0.028 ms for a short sentence, 0.10 ms at 193 characters (p99 0.24 ms). On-device budgets on phone-class hardware will be higher; the design target is under 1 ms.

Training data

Built exclusively from commercially clean components: Tatoeba (CC BY 2.0 FR), Common Voice sentence collections (CC0), Wikidata Lexemes (CC0), Hunspell dictionaries (permissive per-dictionary) and five Universal Dependencies treebanks (CC BY 4.0). Share-alike sources are excluded by policy and by build check: no Wikipedia or Europarl text enters training. Attributions and dataset citations are in THIRD_PARTY_NOTICES.md.