SpeakLoop

Hear the sounds a transcript misses.

An open-source English pronunciation coach for Alexa+, built on MCP. It scores every sound you say, remembers which ones you keep dropping, and drills them tomorrow.

Run it locally Try the card (demo data) GitHub
A card comparing what a typical assistant hears with what SpeakLoop hears, flagging the missing /t/ in 'walked'.

Benchmark

L2-ARCTIC, Vietnamese speakers; the same 100 held-out utterances (3,200 phones, 794 mispronounced). Thresholds were tuned on separate dev speakers only.

SystemPrecisionRecallF1
Whisper-small transcript diff (baseline)0.620.220.32
SpeakLoop phoneme scorer (default)0.590.530.56

The cost: more false alarms on correctly said phones (12.3 % vs 4.5 %). Method and per-error breakdown.

The cards

Screens of the MCP App card rendering demo data. Open any of them live: assessment · every sound · minimal pairs · weak sounds · plan.

Every phone of the sentence, marked heard, changed or missing

Every sound you said, synced to your recording.

Minimal-pair drill card: walk versus walked

A minimal-pair drill for the sound you dropped.

Three weak-sound rings with mastery this week

Weak sounds, tracked across days with FSRS.

Honest status

Alexa+ add-on tooling is partner-only, so SpeakLoop runs in a browser simulator that is a real MCP client. The phone-level scorer is an optional install (open wav2vec2 weights, ~1.2 GB) and is not yet reachable from the simulator's microphone. Details in the README.