hex/ios/HEXKeyboard/ParakeetLiveTranscriber.swift
Ryan Vogel 130835701d
iOS keyboard: Show what I say and Type as I speak (words while you dictate) (#2410)
* iOS keyboard: live words over the keys while dictating (stage 1, two looks)

What: while the keyboard records (in-keyboard dictation, iOS 26+), the capture's
16-bit chunks also feed Apple's SpeechAnalyzer (SpeechTranscriber with volatile and
fast results, DictationTranscriber's progressive preset for its languages), and the
words so far show over the keys: settled words in the keys' ink, the tentative tail
muted, newest line in view, whole lines at the top. Two looks behind a temporary App
Group key (debug.liveDictationLook): dim (keys at 10%) and hide. The preview takes every
touch on the keys, so nothing types while it shows; the recording bar, its cancel and
done, and the globe (drawn over the words where the keyboard has its own) work as before.
The final text still comes from the recording exactly as today; the live recogniser is
stopped when the capture stops (model lingering for the final pass).

Why: Ryan asked for the dictation to show in real time over a hidden or dimmed keyboard.
LiveTranscriptSource is one small protocol so the source can be swapped:
ParakeetLiveTranscriber (the audio so far again about once a second, a 20 s window) is
written but not wired. Never worse than today: no preview when neither switch wants it,
when less than 24 MB is left, when the language's model isn't on the iPhone (no
download), on an error, or before iOS 26. Hooks in KeyboardLocalDictation stay in the
capture's events and phase, away from the transcription call.

Verified: Studio simulator (iPhone 17 Pro iOS 27.2, SE iOS 26.5), real keyboard in a
host app with results recorded from Apple's recogniser on macOS 27.2.1 replayed (the
simulator has no speech models; LiveTranscriptReplay is simulator-only): both looks,
light and dark, upright and turned, a 60 s dictation scrolling, taps on Q and space
while recording type nothing with the preview (and "Q " without it), the globe switches
keyboards on the SE, no model -> no preview and the keys as today. Measured on macOS
with the same source: first words 0.5-1.0 s after the first word ends, word lag median
0.56-0.76 s, 2.1% of shown words rewritten, +0.4-2.6 MB in the feeding process.
code-shape.py passes. No settings UI yet (stage 2).

* iOS keyboard: Show what I say (hide look) and Type as I speak

What: look (b) only. While the keyboard records by itself the keys go and the words
so far take their place under the recording bar; the dim look and its temporary
debug.liveDictationLook key are gone. Two switches in Settings › Dictation on their
own card under the style card, iOS 26 and later only, Simplified Chinese:
- Show what I say (on by default): the preview.
- Type as I speak (off by default, "Words can change until you stop"):
  LiveFieldTyping types the live words through the text document proxy, retyping only
  the changed tail, and keeps exactly what it typed. It edits only while the text
  before the cursor still ends with that (smart quotes allowed for; a host showing only
  the end must show the part that changes); otherwise it stops and leaves the field,
  and the dictation goes in at the cursor as today. When the dictation's text goes in,
  KeyboardLiveTyping.handOver takes the words out first and returns the field as it was,
  so spacing, Undo dictation (the whole dictation), Learn how I type and typing logs see
  exactly what they did before. Cancel, a failure, nothing heard and the keyboard hiding
  take the words out (a held dictation goes in at the next keyboard as before). Never
  in secure, password, one-time code or card number fields, over a selection or a
  two-thumb Dictate's range, or with Translate what I say on (Ryan's call).
- Never worse than today: live words only with 32 MB left once the capture runs (the
  speech model's 24 MB transcription headroom plus the recogniser's share), stopped under
  24 MB or on a memory warning; the preview never shows before a recogniser is running,
  so a dictation without one never flashes it.
- The hand-over: the preview fades with the words it had while the keys and the
  dictation's own text come in; with Parakeet now English's default the field's words
  can differ slightly from the preview's (Apple's), and never flash in the preview.

Why: Ryan picked look (b) and the two switch names at the checkpoint.

Verified: Studio simulator (iPhone 17 Pro, iOS 27.2) with the replayed recogniser
results: the field typed and revised as the words came ("Quick note: Hey, I am running
about 10 minutes late. Can you grab us a table by...") and back to "Quick note:" after
Cancel; Settings › Dictation in English, dark with both on, and Simplified Chinese.
New tests in DictationLearningTests (LiveDictationTests: accumulator, joining, switch
defaults, typing and revising, stopping when the field changes or reports nothing,
private fields). README describes what the keyboard shows while recording.
code-shape.py passes.

* iOS keyboard: the preview draws only the end of a long dictation

What: LiveDictationPreviewView draws its words with one label in a clipping box, moved up by
whole lines once they overflow it, and only about the box's worth of text and three lines more
(from a word's start). It replaces the scrolling UITextView, so the words can no longer be
dragged back.

Why: on the iPhone 17 Pro a 60 s dictation with the preview held up to 27 MB more in the
keyboard (peak 103 MB, against 60 MB with the recogniser on and the preview off); the Studio
simulator showed the same, +24.5 MB.

Verified: Studio simulator, the 60 s replay: +24.5 MB (text view) -> +17.3 (capped text view)
-> +5.7 MB with the label, against +7.0 MB for the same recording without the preview. Test
for the tail in LiveDictationTests.

* iOS keyboard: the preview's tail keeps a word that starts right at the cut

Verified: LiveDictationTests on the Studio simulator.

* iOS keyboard: Type as I speak's words are replaced by the held text after the keyboard hides

What: when the keyboard hides mid-dictation, KeyboardLiveTyping notes the words it typed
(LiveDictationSettings.leftover, in the App Group) as well as taking them out. When the held
dictation goes in (inserted by itself or by Paste dictation), it replaces those words if the
text before the cursor still ends with them (LiveFieldTyping(resuming:in:)); otherwise it goes in
at the cursor as before. A new dictation clears the note.

Why: on the iPhone 17 Pro, pressing Home mid-dictation left the typed words in the field (an
app gone to the background takes no edits) and, back in the app (a new document identifier),
the strip offered Paste dictation, which would have put the dictation after them.

Verified: LiveDictationTests on the Studio simulator (9 tests).

* iOS README: what the keyboard shows while recording, measured on the iPhone 17 Pro

Adds the phone's numbers (first words about 1.2-1.3 s after speech starts; 60 s dictation 60.1 MB
with the preview against 58.1 MB without, peaks 66.5 against 60.1), the hidden-keyboard case for
Type as I speak, and what is verified where and what isn't (VoiceOver on a device, lagging hosts).
2026-11-22 13:37:37 -05:00

122 lines
5.8 KiB
Swift
Raw Permalink Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

import AVFAudio
import Foundation
/// Live words from the keyboard's own speech model, which can't stream: the audio so far is
/// transcribed again about once a second (Parakeet V2 on the Neural Engine reads 15 s in
/// about 0.15 s and a minute in 0.25–0.5 s on an iPhone 17 Pro). Words two passes agree on,
/// short of the last two, are settled; the rest stays tentative. Once a pass covers more than
/// `windowSeconds`, the words ending well before its end are kept as they are and later
/// passes start after them, so a pass never grows past the window. Not yet used:
/// `KeyboardLiveDictation` takes it once the speech model is the keyboard's engine; until
/// then `transcribe` is the model's pass over 16 kHz samples (`KeyboardParakeet`), nil when
/// it can't run one now (another pass or the dictation's own transcription running, short
/// of memory).
@MainActor
final class ParakeetLiveTranscriber: LiveTranscriptSource {
struct Pass {
let text: String
/// Each word's end in the samples passed, when the model gives token timings.
let wordEnds: [Double]?
}
var onUpdate: ((LiveTranscript) -> Void)?
var onFailure: ((String) -> Void)?
private let transcribe: ([Float]) async -> Pass?
private let inputFormat: AVAudioFormat
private let modelFormat: AVAudioFormat
private let converter: AVAudioConverter?
private var samples: [Float] = []
/// Where the window starts in `samples`; the words before it are `frozen`.
private var windowStart = 0
private var frozen = ""
private var previousWords: [String] = []
private var settledCount = 0
private var samplesAtLastPass = 0
private var pass: Task<Void, Never>?
private var stopped = false
static let sampleRate: Double = 16_000
/// New audio before another pass.
static let interval: Double = 1.0
static let windowSeconds: Double = 20
/// Words ending this close to a pass's end are never frozen.
static let keepTentative: Double = 5
init?(sampleRate: Double, transcribe: @escaping ([Float]) async -> Pass?) {
guard let input = AVAudioFormat(commonFormat: .pcmFormatInt16, sampleRate: sampleRate, channels: 1, interleaved: true),
let model = AVAudioFormat(commonFormat: .pcmFormatFloat32, sampleRate: Self.sampleRate, channels: 1, interleaved: false) else {
return nil
}
inputFormat = input
modelFormat = model
converter = AVAudioConverter(from: input, to: model)
self.transcribe = transcribe
}
func append(_ pcm: Data) {
guard !stopped, let converter else { return }
let frames = pcm.count / 2
guard frames > 0, let source = AVAudioPCMBuffer(pcmFormat: inputFormat, frameCapacity: AVAudioFrameCount(frames)) else { return }
source.frameLength = AVAudioFrameCount(frames)
pcm.withUnsafeBytes { raw in
guard let base = raw.baseAddress, let target = source.int16ChannelData?[0] else { return }
memcpy(target, base, frames * 2)
}
let capacity = AVAudioFrameCount((Double(frames) * Self.sampleRate / inputFormat.sampleRate).rounded(.up) + 32)
guard let converted = AVAudioPCMBuffer(pcmFormat: modelFormat, frameCapacity: capacity) else { return }
let once = OneShotInput(source)
var error: NSError?
converter.convert(to: converted, error: &error) { _, status in once.next(status) }
guard error == nil, let channel = converted.floatChannelData?[0] else { return }
samples.append(contentsOf: UnsafeBufferPointer(start: channel, count: Int(converted.frameLength)))
runPassIfDue()
}
func stop() {
stopped = true
onUpdate = nil
pass?.cancel()
pass = nil
samples = []
}
private func runPassIfDue() {
guard pass == nil, Double(samples.count - samplesAtLastPass) >= Self.interval * Self.sampleRate else { return }
samplesAtLastPass = samples.count
let window = Array(samples[windowStart...])
pass = Task { [weak self] in
let result = await self?.transcribe(window)
guard let self, !self.stopped else { return }
self.pass = nil
if let result { self.take(result, windowSamples: window.count) }
self.runPassIfDue()
}
}
private func take(_ result: Pass, windowSamples: Int) {
let words = result.text.split(whereSeparator: \.isWhitespace).map(String.init)
var agreed = 0
while agreed < min(words.count, previousWords.count), words[agreed] == previousWords[agreed] { agreed += 1 }
settledCount = min(words.count, max(settledCount, agreed - 2))
previousWords = words
let settled = (frozen.isEmpty ? [] : [frozen]) + words.prefix(settledCount)
let tentative = words.dropFirst(settledCount)
onUpdate?(LiveTranscript(settled: settled.joined(separator: " "), tentative: tentative.joined(separator: " ")))
freezeIfLong(words: words, ends: result.wordEnds, windowSamples: windowSamples)
}
/// Past the window, the settled words that end `keepTentative` before the pass's end are
/// kept and the next window starts after the last of them.
private func freezeIfLong(words: [String], ends: [Double]?, windowSamples: Int) {
let seconds = Double(windowSamples) / Self.sampleRate
guard seconds > Self.windowSeconds, let ends, ends.count == words.count else { return }
var count = 0
while count < settledCount, ends[count] < seconds - Self.keepTentative { count += 1 }
guard count > 0 else { return }
frozen = ([frozen] + words.prefix(count)).filter { !$0.isEmpty }.joined(separator: " ")
windowStart += Int(ends[count - 1] * Self.sampleRate)
previousWords = Array(previousWords.dropFirst(count))
settledCount -= count
}
}