How does SoraSpeak work?

Four steps from sign-up to your first scored conversation

Pick a teacher, choose a topic, speak naturally, see your scores per phoneme. Thirty seconds to start your first session — no card, no setup, no install.

01

Choose your AI teacher

Pick between two teachers. Aoi speaks slowly and waits, for everyday 日常会話. Ren corrects the politeness level itself, for business Japanese and 敬語.

  • Everyday 日常会話 and business 敬語 registers
  • Beginner-friendly to advanced
  • Switch anytime — no commitment

Aoi

Everyday

日常会話

Ren

Business

敬語
02

Pick a topic

Choose what you want to talk about. Daily conversations, travel scenarios, business meetings, job interviews — real situations you'll actually face.

  • 6 topic categories
  • New conversation every time
  • Matched to your interests
Daily Conversation
Travel Japanese
Business Meeting
Job Interview
Casual Chat
Debate & Opinion
03

Have a real conversation

Speak naturally into your mic. The AI responds to what you actually say — no scripts, no fill-in-the-blank. When something sounds off, you get an instant tip to fix it.

  • Free-flowing, not scripted
  • Real-time feedback on mistakes
  • Think in Japanese, not translate
Aoi
週末はいつも何をしますか?
家族と公園に行きます。
✓ Great sentence structure! Very natural.
いいですね!公園で何をするのが好きですか? 🌳
04

Watch yourself improve

Track your speaking confidence over time. Weekly charts show exactly where you're getting better and where to focus next. Consistency beats intensity.

  • Weekly progress charts
  • Session-by-session tracking
  • AI insights on your weak points
Speaking confidence
85↑ 43 pts since week 1
W1
W2
W3
W4
W5
W6
W7
W8

Methodology

How SoraSpeak scoring actually works

Most apps tell you "AI scores your pronunciation" without saying what AI, what scoring, or how. Here is the honest stack behind every spoken response — written so a technical reader can verify the claims.

Step 1 — Speech-to-text

Production-grade speech recognition

Sub-second

transcription latency on short utterances

Every audio chunk is transcribed by an industry-leading multilingual automatic speech recognition (ASR) engine. The engine was trained on hundreds of thousands of hours of speech across many languages, accents, and noise conditions — which is why it handles learners' Japanese across different proficiency levels and microphones.

Inference runs on a low-latency platform that delivers transcripts in well under a second on short utterances. That is what keeps the conversational loop feeling natural rather than transactional. Audio is streamed in roughly half-second intervals.

Step 2 — Conversational reply

Purpose-tuned conversational AI

2 teachers

register + level + pedagogy per teacher

The transcript joins a rolling conversation context and is sent to a fast conversational language model. Each teacher has a system prompt that encodes their persona, target register, proficiency band, and pedagogical strategy — gentle correction, scaffolded follow-ups, and vocabulary expansion relative to the learner's level.

We picked this model class because the conversational loop is cost-and-latency sensitive. A larger model would respond a few hundred milliseconds slower, which breaks the rhythm of real conversation.

Step 3 — 4-axis scoring

What we actually grade

Pronunciation

How closely each word's phonemes match the selected target Japanese sound and rhythm. Errors are surfaced with concrete guidance — for example, a long vowel shortened or a mora dropped — and a native audio sample is provided for comparison.

Grammar

Sentence-level grammaticality, focused on errors that block comprehension rather than minor stylistic variation. Native speakers make "errors" constantly without losing meaning; we grade against communication, not textbooks.

Fluency

Pace, hesitations, filler-word density, mid-sentence restarts. Distinct from accuracy — many learners are accurate but stilted, or fluent but inaccurate. Both axes matter for how natural you sound.

Vocabulary

Lexical range relative to your Japanese level, with suggestions for higher-register alternatives where appropriate. Encourages variety without pushing rare words that would sound forced.

Response feel

Why the loop feels conversational

~1-2 s

end-to-end response time

Real human conversation is sub-second turn-taking. If the loop takes four or five seconds, learners stop and the practice loses its rhythm. SoraSpeak targets the under-two-second band — the threshold above which conversation stops feeling like conversation and starts feeling like a chatbot.

Across a typical home connection, the time from when you stop speaking to when the teacher starts replying is in the one-to-two second range. That budget is what made several engineering choices necessary — model selection, streaming protocols, and avatar lip-sync timing all serve that single number.

Numbers so far

What learners are actually doing

Approximate figures from SoraSpeak's early cohort. Numbers refreshed manually each quarter — we publish round figures, not vanity precision.

+15 pts

avg. score gain

after 30 days of daily practice (10+ min/day cohort, internal data)

~12 min

avg. session

typical conversation length for engaged learners

2 teachers

speaking registers

Aoi for everyday conversation, Ren for business Japanese and keigo

320 topics

conversation prompts

across 10 categories from daily life to technology

Numbers above are rounded approximations from SoraSpeak's internal analytics, intended as honest indicators rather than exact metrics. Individual results vary widely with practice consistency.

What the research says

Speaking-first is not a new idea

The case for speaking practice over grammar drilling is over a century old. Modern AI tools are new, but the pedagogy they implement is built on a long line of research and teaching practice. SoraSpeak didn't invent this — we wrote software for what these people already proved.

The first requisite is a sound knowledge of phonetics. Without it, the pupil's ear remains insensitive to differences in pronunciation, and his organs of speech are not trained to make them.
Henry Sweet (1845–1912)English phonetician, author of The Practical Study of Languages (1899). Foundational work on phonetic teaching.
The teaching of pronunciation must precede everything else, even the teaching of vocabulary, because nothing depresses a learner more than the sense that he cannot make himself understood.
Otto Jespersen (1860–1943)Danish linguist, How to Teach a Foreign Language (1904). Influential proponent of the Direct Method.
Anyone who is willing to take the trouble can learn to pronounce a foreign language reasonably well, provided he has access to a good model and a method of practising systematically.
Daniel Jones (1881–1967)Phonetician and creator of the cardinal vowel system. An Outline of English Phonetics (1918).
We acquire language in only one way: when we understand messages. We call this comprehensible input. The acquisition device fires automatically when the input is understood.
Stephen Krashen (b. 1941)Linguist, Professor Emeritus at USC. Input Hypothesis (1985); foundational theory in modern second-language acquisition.
The student should hear, speak, read, and write in the foreign language, in that order, just as a child learns its native tongue.
Maximilian Berlitz (1852–1921)Founder of the Berlitz language schools and codifier of the Direct Method. The principle still underlies most modern speaking-first pedagogy, including SoraSpeak.

These are not slogans we invented. They are summaries of arguments these researchers made in print, citable in any university library. SoraSpeak is the modern toolchain for ideas that were already correct a hundred years ago.

Start in 30 seconds

No signup required. Just tap the mic and speak.