Big Talk
A speaking gym for iPhone that scores every take on the phone. Formerly SpeakUp.
The floor is −61 dB, so a frame counts as voice above −51 dB. The 0.70 s after um counts as a pause.
Drag the pause or turn up the room. The audio and the transcriber's ranges are made up; the refiner and the 0.4 s pause rule are the app's.
Record about a minute of speech and get a 0–100 score broken into pace, fillers, pauses, clarity, vocal variety and more, with the transcript marked up word by word and concrete coaching. Transcription and scoring run entirely on device: no account, and no audio is sent anywhere to be analyzed.
What it is
Capable people get underestimated when their delivery doesn't carry their ability. Big Talk is a private place to rehearse the first take: a prompt, a talk you're memorizing, an interview answer. You speak, and a minute later you know what to fix.
Most speech coaches send your voice to a server. This one runs Apple's on-device speech model, its own signal processing, and a scoring engine on the phone, so it works offline and your voice is never sent off to be scored. iCloud sync of recordings is opt-in. An optional on-device language model rewrites the coaching in plain English.
How it's put together
Trade-offs
Apple's on-device transcriber, not Whisper or a cloud API. Audio is never sent off to be transcribed, there is no model to ship, and the system speech process runs outside the app's memory budget. The downside: iOS 26 only, device only (the simulator can't transcribe), and less control over the model.
Recover pauses from the raw audio. Transcribers tile words edge to edge, which erases silence. Re-measuring each word against the noise floor gives real pause, filler-context and fluency data. The downside: An extra signal-processing pass per take.
The LLM can nudge the score, never decide it. Small on-device models are noisy. Rule-based scores are the anchor; model input is blended in, clamped (±14 overall with Apple Intelligence, ±10 with the local model), and the gibberish gate runs again after. The downside: The model can't fix a large scoring mistake on its own.
Next: TokenuinelyA semantic, graph-aware index of your repo, served to AI coding agents over MCP.