Apple SpeechAnalyzer, up close.

What the two classes do

• SpeechAnalyzer holds the modules, accepts audio and controls the session, one input sequence at a time. SpeechTranscriber is the transcription module, which Apple calls appropriate for normal conversation and general purposes.
• Audio goes in as an async sequence of AnalyzerInput values, and results come out of the transcriber's results sequence, which you read in a separate task. Each result carries the time range of the audio it covers.
• With the volatileResults reporting option the transcriber sends tentative text for a range of audio, refines it, then closes the range with one final result. Without it you get only final results.
• DictationTranscriber is the fallback for devices SpeechTranscriber does not support. Apple says it uses the same models as SFSpeechRecognizer set to on-device work. It is also new in the 26 releases, so it reaches older devices, not older systems.

Where the model comes from

You do not ship it. Apple's AssetInventory page says these assets are machine learning models downloaded from Apple's servers and managed by the system: once one is installed the system retains it, updates it and shares it with other apps. The WWDC session adds that it sits in system storage and runs outside your app's memory space.

There is no language list in the framework reference, so do not hard-code one. Read supportedLocales for what can be installed, installedLocales for what is already there, and supportedLocale(equivalentTo:) to map a locale such as ja onto one the model has. Apple warns that this last call can return a near equivalent with a different region.

How Fraze drives it

• Apple documents contextualStrings as words or phrases that should be recognized even if they are not in the system vocabulary. Keep each to one or two words, and under 100 across all tags.
• Apple's text describes that property for DictationTranscriber and does not say what SpeechTranscriber does with it, so Fraze treats the terms as a hint and keeps going if the call fails.
• When standard input closes, the helper ends the input stream and calls finalizeAndFinishThroughEndOfInput. The last volatile lines become final and the results loop ends.

Fraze runs SpeechAnalyzer in a small Swift helper. The app writes 16 kHz mono 16-bit PCM to its standard input and reads one JSON object per line back, so a fault in the framework does not take the app down. The transcription path, trimmed:

import Speech

// Inside an async function. Pick a locale the model actually has; this can
// return a near equivalent with a different region.
guard let locale = await SpeechTranscriber.supportedLocale(equivalentTo: Locale(identifier: "ja"))
else { return }

let transcriber = SpeechTranscriber(
  locale: locale,
  transcriptionOptions: [],
  reportingOptions: [.volatileResults],
  attributeOptions: [])

// The model is a system asset. Download it once; the system keeps and updates it.
if await AssetInventory.status(forModules: [transcriber]) != .installed,
   let request = try await AssetInventory.assetInstallationRequest(supporting: [transcriber]) {
  try await request.downloadAndInstall()
}

// The analyzer does not resample. Convert your audio to this format yourself.
guard let format = await SpeechAnalyzer.bestAvailableAudioFormat(compatibleWith: [transcriber])
else { return }
let analyzer = SpeechAnalyzer(modules: [transcriber])

// Custom vocabulary: names and terms to bias recognition toward. A hint, so
// a failure here is ignored.
let context = AnalysisContext()
context.contextualStrings[.general] = ["Fraze", "Parakeet"]
try? await analyzer.setContext(context)

// inputSequence is your AsyncStream<AnalyzerInput>: AnalyzerInput(buffer:) values
// whose AVAudioPCMBuffer is already in `format`.
try await analyzer.start(inputSequence: inputSequence)

for try await result in transcriber.results {
  let text = String(result.text.characters)
  print(result.isFinal ? "final: \(text)" : "volatile: \(text)")
}

Translation, and why only the low-latency strategy

The Translation framework is separate. TranslationSession has existed since macOS 15 and iOS 18, but asking for a strategy is new: Apple lists TranslationSession.Strategy as iOS 26.4 and macOS 26.4. The docs describe lowLatency as fast translation with traditional models, for cases such as translating audio in real time, with languages downloaded before use. The other strategy, highFidelity, uses Apple Intelligence models and may take longer.

Apple's page for the class says all translations using TranslationSession are processed on the user's device. Fraze asks for lowLatency and nothing else, for two reasons. It is the one Apple points at real-time audio. And its state can be checked: Fraze asks LanguageAvailability, with the same strategy, whether a pair is installed, and does not use Apple translation for the pair otherwise.

The second reason comes from an observation of our own. Without a strategy set, a session returned translations for pairs whose language packs we had not added in System Settings. Apple's highFidelity page gives a possible cause: those models are already downloaded when Apple Intelligence is on, so no language download is needed. We did not test further.

How accurate is it

• Spanish: Parakeet 6.05, Apple 7.04, hosted Whisper 5.68, Soniox 5.04.
• French: Parakeet 7.83, Apple 8.42, hosted Whisper 6.82, Soniox 5.72.
• Russian: Parakeet 9.18, hosted Whisper 7.48, Soniox 6.52. The Apple engine had no Russian in our run.
• Korean: Apple 15.61, hosted Whisper 13.66, Soniox 9.62. Parakeet v3 has no Korean.
• Chinese: Apple 6.76, hosted Whisper 6.01, Soniox 4.71. Parakeet v3 has no Chinese.

Word error rate in percent, lower is better, from our own runs on an Apple silicon Mac. Japanese and Chinese are character error rate. The per-language numbers are FLEURS read speech, 200 clips each, measured on 2026-08-30.

Whisper large-v3-turbo appears twice. One row ran on-device through WhisperKit, a Core ML port, not OpenAI's reference PyTorch build. The other is a hosted copy, the only one we ran on FLEURS. Parakeet ran through parakeet-mlx, not the int8 sherpa-onnx build Fraze ships.

What that means, and the limits

• Apple platforms only, and no way to run the model outside Apple's frameworks.
• The 26 releases only, and not watchOS. DictationTranscriber is new in 26 as well, so on older systems the choice is SFSpeechRecognizer.
• The WWDC session mentions hardware requirements, so check SpeechTranscriber.isAvailable first.
• The model is closed. Apple publishes no weights, and you cannot pin a version: Apple says the system updates the assets automatically.
• Choosing a translation strategy needs macOS 26.4 or iOS 26.4.

On clean read English the top is a tie. Apple's 1.82 has a 95 percent interval of 1.66 to 1.99, which takes in Whisper turbo. LibriSpeech is in Parakeet's training data, and several cloud engines scored lower than all of these.

Conference calls are closer to what a meeting sounds like. There Apple scored 12.03 (11.52 to 12.56) against Parakeet v3's 11.20 (10.71 to 11.74). The intervals overlap, but the benchmark's paired test on the same clips puts Apple 0.5 to 1.2 points behind Parakeet, and level with Whisper through WhisperKit at 12.35 (11.87 to 12.86). The repository says the reference PyTorch Whisper turbo scores about 11.1 on this set, level with Parakeet.

Outside English the cloud engine led Apple by two to three points, and by six on Korean, with hosted Whisper in between. Apple was about a point behind Parakeet on German, Spanish and French. At 200 clips the intervals run about a point either side, so read small gaps as ties. Apple and Parakeet both ran at about 0.02 times real time. The repository does not rank engines on speed.

The Fraze row shows which engine the app runs: Apple's on macOS 26 where it has the language, Parakeet otherwise. What the Apple engine adds for Fraze is languages: Japanese, Korean and Chinese, which Parakeet v3 does not have.

Questions