voice-first · hands-free · sub-1.5s latency

Learn Spanishby talking,not tapping.

A real-time conversational tutor. Put your phone down. Speak. The agent teaches lessons, runs quizzes, roleplays cafés in Madrid, and answers your doubts mid-sentence — all by voice.

Start talkingSee what it can do
4
modes
teach · quiz · convo · doubt
6
lessons
A1 → A2 curriculum
<1s
P50 latency
end-to-end measured
Groq · 200 tok/s
Pipecat 1.2
§01 · live demo

Click the orb. Talk.

Mic permission needed. Headphones recommended (kills bot-echo feedback). All processing local except STT / LLM / TTS calls.

Idle
·offline
Click the orb to begin

Try saying

"Teach me greetings"
"Quiz me on yesterday"
"Let's roleplay a café"
"I have a doubt"

Live Transcript

0 turns
transcript will appear as you speak…
§02 · capabilities

Built for the parts
that actually matter.

Every feature was picked against the assignment's evaluation matrix — system design, voice UX, pedagogy, reliability, observability — with the bonus stretch goals layered on top.

Real barge-in

Silero VAD + Pipecat interruption frames cut bot mid-word in <250 ms. Talk over it freely.

measured

Sub-1-second response

P50 end-to-end ~880 ms on free-tier providers. AssemblyAI streaming → Groq → ElevenLabs Turbo.

Structured teaching

FSM-enforced lesson flow: objective → explain → example → practice → check. No rambling.

Semantic quiz grading

Paraphrases accepted. Two-tier: deterministic Jaccard first, LLM only for ambiguous middle.

Doubt resume

Mid-lesson "wait, why is it me llamo?" → answers in English, returns exactly where you left off.

bonus

FSRS-lite scheduling

Each quiz answer feeds spaced-repetition. Weak words come back tomorrow, mastered ones in a week.

bonus

Prosody engagement

Real-time RMS + pace + pause analysis. Bot detects when you sound tired and slows down.

Phoneme-aware feedback

Not "good job" — specific tips like "the rr in perro is rolled, try humming and tapping".

Per-turn observability

Every turn logs STT / LLM / TTS / total latency to JSONL. Answer 'why did turn 14 take 2.3s'.

bonus

Multi-persona handoff

Teacher / Examiner / Companion — same LLM, three styles, switched by FSM mode.

§03 · curriculum

Six hand-authored lessons.
A1 to A2.

Hand-authored over LLM-generated so a reviewer can spot-check grammar. Adding a seventh lesson is a JSON edit — no rebuild.

01A1

Greetings & Introductions

Greet, introduce, ask how someone is.

10 vocab · 5 min
02A1

Numbers 1 to 20

Count and use numbers in phrases.

20 vocab · 6 min
03A2

Ordering Food

Order politely, ask for the bill.

12 vocab · 7 min
04A1

Family Members

Talk about parents, siblings, relatives.

11 vocab · 6 min
05A2

Days of the Week & Time

Name days, tell time at the hour.

12 vocab · 7 min
06A2

Asking for Directions

Ask where places are, understand directions.

12 vocab · 8 min
§04 · stack & latency

Open tools,
measured performance.

STT
ms p50
p95 ms · target < 200 ms
LLM TTFT
ms p50
p95 ms · target < 500 ms
TTS first
ms p50
p95 ms · target < 450 ms
E2E total
ms p50
p95 ms · target < 1500 ms