Sonja PhraseFlow

Author: Sonja Sahebzad
Utrecht, the Netherlands | September 27, 2026

Your next word,
in your language.

Twelve selectable languages.
A compact R app with measured results and a clear user guide.

Open the illustrated user guide

Sonja Projects · Version 1.4
English is unchanged. Additional languages are experimental.
Open the live app

The prediction model

A separate model for each language

English, German, Finnish and Russian use course-corpus text. Eight added languages use Tatoeba example sentences.

Recent words guide the next word

A five-gram model uses up to four recent units. Kneser-Ney smoothing blends longer and shorter patterns when evidence is sparse.

Text processing follows the script

Chinese uses ICU word segmentation. Korean retains space-delimited units. Accents, Cyrillic and Hindi vowel signs are preserved.

12selectable languagesCPUnative R inference

Lazy loading and a bounded cache.
No external prediction API.

Methods: Chen & Goodman; ICU. Additional data: Tatoeba contributors, CC BY 2.0 FR.

Using the app and the guide

  1. Choose a language and enter a phrase.
  2. Finish a word, then pause or select Predict next word.
  3. Click a suggestion to keep writing.

Chinese does not require spaces. Choose 5, 10 or 20 words. Extra buttons offer short endings from matching training phrases.

Open the illustrated user guide

The User guide / Handleiding button explains both features. Short endings remain experimental; the accuracy charts measure the word model.

Sonja PhraseFlow language selector, phrase input and next-word suggestions.

English performance

English first suggestion accuracy 17.0%, top three 28.6%, with 95% intervals.

900 unseen examples

First: 153/900
Within three: 257/900

Whiskers show 95% sampling intervals. They do not express certainty for one suggestion.

Fast local lookup

Archived median: 0.58 ms for three suggestions. Network and display add time.

Archived English evaluation; test examples were not reopened. All 900 received a suggestion. This is availability, not 100% accuracy. Individual confidence is not calibrated.

Additional language results

Initial accuracy in eight additional languages, with first-suggestion and top-three scores and 95% intervals.

600 cases per language

Held-out Tatoeba sentences. Settings were fixed before evaluation.

Different data, training sizes and word units: these are separate results, not a language ranking.

Next improvement

Better representative data, native-speaker review and a compact model using more sentence context.

Lines show Wilson 95% sampling intervals. Similar sentence templates may remain; everyday-message performance is untested. Full results, earlier prototypes, timing and instructions · Data credits. Sonja Sahebzad · Utrecht, the Netherlands.