Spee. Get the code

Bee listens.
Spee tells me what to change.
I do the talking.

Spee learns how I usually speak from my real conversations in Bee, then tells me what to change: on my wrist, in a recap, in a drill with Alexa, and on my TV.

See how it works Watch the demo

How it works

1

Bee hears my real conversations.

Bee records on my Apple Watch and tags the lines that are mine. Its live tag named me on 8 of 9 lines when I measured it. Other voices are counted as words, never kept.

2

Spee counts them against my usual.

Fixed word lists count unsure words, filler “like”, vague answers and starting over, per 100 words, against how I usually speak in that setting. No model does the counting.

3

I get one thing to change.

A streak buzzes my wrist about 20 seconds after the line. After the conversation, a recap names what ran high and gives one of my lines back, said plainly.

English is my second language. Nobody stops a stand-up to point out a filler word, so I built something that does it quietly. It counts four habits.

I guessmaybekind ofI think

Practised. The habit I most want gone.

// prereg/transfer.json, October 16
"drilled": [
  { "habit": "hedges", … },

The whole loop.

One hub, four ways back to me. Pick any step.

Checked by Bee

Spee hub

Fixed word lists count my habits against my own usual.

Unsure words, filler “like”, vague answers and starting over, per 100 words, compared with how I usually speak in that setting: a meeting, work talk or a casual chat. Code decides; no model does the counting or the grading.

One stand-up, step by step.

Stand-up · work talkSynthetic

  1. 0:41So the export is, I guess, nearly done.
  2. Someone else talks: counted, never kept.
  3. 1:05The CSV part is kind of working now.
  4. 1:21We could maybe test it tomorrow.
  5. 1:34And I think we ship it on Friday.
  6. 1:51My wrist buzzes
  7. 3:4090 s of quiet: the conversation ends

LiveCoaching on

2unsure words so far

Nothing unusual yet, so Spee stays quiet.

Every screen, from one hub.

  • Spee on the TV, Live: the practice line on the left; this conversation on the right with a wrist nudge, 4 unsure words in 2 minutes, and its hops: said, Bee at 5.9 s, wrist alarm at 17.4 s (synthetic demo data)

    Live. The drill on the left, the conversation on the right. The nudge shows each hop: me, Bee, the wrist alarm.

  • Spee on the TV, Today: 4 lines to practise, my week per 100 words, and the before/after check (synthetic demo data)

    Today. What to practise, my week per 100 words, and whether it’s working.

  • The Spee recap on the TV: unsure words ran high, 4.2 per 100 words against a usual 1.4 at work, with one line said plainly (synthetic demo data)

    Recap. What ran high against my usual at work, and one of my lines said plainly.

  • Spee Progress on the TV: unsure words and filler like dropped after the first practice; starting over, not practised, did not change (synthetic demo data)

    Progress. Before and after my first practice, next to the habit I didn’t practise.

  • The Alexa+ simulator: quiz me, the line said back, then how did I do?, answered Passed once Bee heard it (synthetic demo data)

    Alexa. In the browser simulator: “quiz me”, say the line, “how did I do?”. The pass waits for Bee.

Four surfaces, one hub.

  • Apple Watch with Bee / Amazon Bee. Records, and buzzes with a nudge or a recap.

  • Alexa. “Quiz me”, “how did I do?”, “how am I doing?”, “did it transfer?”.

  • Fire TV. A Vega app for my week and the live drill.

  • Phone. Drills and practice meters. No accounts.

Other voices are counted, never kept.

Mine onlyOnly lines Bee tags as me are kept

Other voicesA word count, never text on a screen or disk

AudioBee stores none

What I measured.

  • 5.4 s

    median time for Bee to tag a line after I stop speaking (4.3 to 14.7 s, October 16)

  • ~20 s

    from the line to the buzz on my wrist: Bee’s tag, the to-do call and a 10 s alarm lead

  • 8 of 9

    lines where Bee’s live tag named me, so the coach doesn’t guess who spoke

  • 400

    words, at least, on each side of my first practice before the before/after check gives a verdict

Questions.

How fast is the nudge?

Bee tags a line about 5 s after I stop speaking (median 5.4 s, 4.3 to 14.7 s, measured on 2026-10-16), so a wrist nudge lands about 20 s after the line that triggered it.

Why does a drill verdict take a turn?

A drill line can take up to about 20 s, because Bee may merge it with my next words, and Alexa must answer within 8 s, so the verdict comes on the next turn.

When does my Watch buzz?

Only while Bee is recording and the iPhone is locked. Otherwise the alarm shows on the iPhone. Nudges are off unless I turn them on, and never from 22:00 to 08:00.

Why a custom skill and not Alexa+?

Alexa+ add-ons are partner-gated, so Alexa is a classic custom skill plus an MCP server any assistant can use. The browser simulator is that MCP client.

Does it hear my tone of voice?

Bee stores no audio, so tone and pitch only exist in phone practice. Bee coaching works from words and timing.

Is a wearable speech coach new?

No. A 2025 study (arXiv 2507.04238) reports that near-real-time feedback and after-the-fact review both reduced unwanted words. What I add is Bee as the grader, drills on my own lines, and a check against a habit I don’t practise.

What does the demo show?

The demo replays Bee events with synthetic data, and every screen says so. With a Bee account, the same hub runs on real data.

The demo.

Under three minutes, on the real devices.

The video goes here once it’s on YouTube.

Until then, step through the stand-up or run the demo below.