monkeyspeak final devlog before ship ( hopefully )

---

tl;dr

• so brave and edge broke everything ( ye read what broke for more info )
• deepgram mode now transcribes live text again (words dissolve, wpm moves, momentum still reacts to your voice)
• brave and edge no longer hang for 25 seconds then fake a mic error
• production uses a render websocket proxy + vercel env, same path that worked locally
• github: nothariharan/monkeyspeak (commit 095b731 and earlier speech routing work on main)
• live: monkeyspeak-delta.vercel.app

---

what i changed

speech routing (back to something that made sense)

mode behavior browser web speech api only. no deepgram hijack on brave/edge mount. deepgram try deepgram first (proxy → bridge fallback on chrome only). if that fails, fall back to web speech with a clear error.

the ui now shows errors for the provider that actually failed, not a generic “mic blocked” when stt died for other reasons.

deepgram client hardening

• prefer var name ur wish websocket proxy over the vercel http bridge
• brave/edge without a reachable proxy: fail fast with a useful message (no 25s timeout)
• utterance_end_ms locked to 1000 everywhere (deepgram rejects live ws with 400 below that)
• bridge watchdog clears on BridgeReady, not on first random chunk
• server bridge waits for upstream deepgram before closing the socket

the transcript fix (the big one)

• client parses deepgram json whether it arrives as a string, blob, or arraybuffer
• render proxy forwards deepgram replies as utf-8 text frames instead of opaque binary

infra

• backend/ express + ws proxy deployable on render (render.yaml included)

• set up env with ur var name at vercel

• redeployed frontend via vercel cli after env update

small polish

• clearer vad fallback logs (worker load fail vs no voice detected vs timeout)
• config bar hint on brave/edge when browser mode is selected
• momentum sprites + gsap monkey states (from earlier in the sprint)

---

what broke (and why it looked cursed)

1. brave and edge block or mishandle the vercel http audio bridge (duplex fetch upload), so deepgram connections hung ~25s and never returned transcripts.

2. those browsers also cannot open an authenticated websocket straight to api.deepgram.com, so they need the render ws proxy (wss://…/api/deepgram/proxy) instead of the chrome-friendly paths.

3. even after the proxy worked, deepgram sent json in binary ws frames and the client ignored anything that was not a string, so words never updated until we parsed blob/arraybuffer payloads.

---

what’s next (maybe)

• vendor ort wasm for vad so the worker stops complaining
• gsap scale split for clean console
• render keep-alive or paid instance if cold starts hurt demos
• preview env on vercel for pr deployments (production is wired today)

---

sorry if i yapped a lot and the fixes where actually slighly more techy stuff so i didnt wanna yap abt that as wel so yeah if u want to know lmk in replies :)

reworked the ui and got monkeyspeak ready for v0 ship launch :))

it’s basically monkeytype but for your voice — read a prompt out loud, get scored on speed + clarity.

features:

• speed mode: timed speaking tests (15s / 30s / 60s / 120s) with live wpm + word accuracy
• clarity mode: paste your transcript, get a word-by-word diff + letter grade
• click the monkey mascot to start (no separate mic button anymore lol)
• animated monkey companion that reacts to your speaking energy mid-test
• words dissolve on screen as you say them correctly
• personal bests saved locally
• browser speech api works out of the box, optional deepgram for better stt
• themes, accent colors, custom fonts
• keyboard shortcuts (enter to start, tab to reset, escape to stop)

built with next.js, typescript, gsap, zustand.

repo: github.com/nothariharan/monkeyspeak

made the ui more clean still settling the latency here and there

current new approach i am building around is using the deepgram api key for initial purpose later trying to replicate what they do

open to questions or recommendations for the whole stt etc.

cooking up 🚀🔥

soon making it public just lots of latency issue in speaking to text conversion who knew it would be a pain but yeah let me know if u guys think of an alternative for it or anything at all

thanks for reading :)

going with a much better ui soon launching :) 🚀
for yall to try it

built the first iteration of the website !

got lots of customizing options now i am working more towards reducing the latency between what u speak and what is being transcribed

give me your opinions and recommendations :)