AnySpeechRecognizer, announced on October 29, 2026, gives Apple-app developers a choice of local speech models and a way to separate speakers in transcripts. It is a software component for building recording and captioning tools, not a finished meeting app.
The available version 0.1.0 Swift package offers Apple, Parakeet and Qwen recognition options. For interviews and small meetings, the key choice is whether the app needs timed words and speaker labels, or Qwen’s additional language coverage.
Apple’s SpeechAnalyzer already transcribes on the device after downloading its models. The reason to consider this package is a different recognition engine or its added speaker workflow—not local processing alone.
Choose the output before the model
What the app needs | Starting point | Main consideration |
|---|---|---|
Keep existing on-device transcription | Stay with it if its supported languages and devices meet the task. | |
Live captions with timed words | Both provide partial text that updates as speech arrives. Unified is an English-only alternative. | |
Recorded Mandarin, Japanese, Korean or Hindi | Completed segments, without live partial text or word timestamps. | |
Speaker-labelled interviews or small meetings | Requires word timestamps; tracks up to four speakers. |
Check the actual language, not just the count. Parakeet v3’s 25-language list includes Ukrainian and Maltese; Qwen’s 30-language list does not. Qwen includes Hindi and Japanese, which Parakeet v3 lacks. Listed coverage is not proof of equal accuracy across languages or accents.
Speaker labels need more than transcription
Speaker diarization means separating who spoke when; word timestamps locate each word in the recording. The package’s word-to-speaker assignment matches timed words against a speaker timeline and skips untimed text. This Qwen integration has no word timestamps, so adding Sortformer does not produce a ready-made speaker-labelled Qwen transcript. Its completed audio segments are not a substitute for live captions.
The documented speaker labels are numbered slots within a session, not people’s identities. Detecting overlapping voices also cannot recover words the transcriber missed. The maintainer notes that transcription commonly captures the louder voice and can omit brief interjections.
NVIDIA’s four-speaker Sortformer v2.1 model card warns of weaker performance with five or more speakers, non-English speech, noise and very long recordings. A language supported by the transcriber does not guarantee dependable speaker separation.
In Mattt’s reported M4 Max Mac Studio tests, every tested engine had a word error rate above 23% on overlapping AMI meeting recordings. That measure counts substituted, missing and extra words against a reference transcript; it is not a percentage of sentences judged wrong. Speaker-labelled meeting notes should therefore remain editable.
Budget downloads separately from memory
An illustrative live Parakeet v3 setup with speaker labels totals about 1.47 GB: 632 MB for the model + 591 MB for the live encoder + 246 MB for Sortformer = 1,469 MB. This adds rounded published download sizes in decimal units; it excludes app files and compilation artifacts and is neither measured installed size nor RAM use.
The Qwen conversions list 0.94 GB of files for 0.6B and 2.36 GB for 1.7B. Their author reports roughly 1.7 GB and 4.4 GB of peak transcription memory, respectively, on an M4 Max Mac—not on an iPhone. The larger model’s download is about 2.5 times the smaller one, calculated as 2.36 ÷ 0.94.
The package code carries Apache 2.0, but downloaded models have separate licences. Check the chosen model’s terms and notices rather than assuming the wrapper’s licence covers its assets.
Declared iPhone support is not tested readiness
The package manifest declares Swift 6.2, macOS 26 and iOS 26 minimums. Parakeet and Sortformer require Apple silicon. The maintainer’s testing caveat is narrower: macOS 27.0.1 on an M4 Max Mac Studio, with physical iPhones, the OS 26 runtime and the minimum Swift toolchain untested. Treat iPhone deployment as work to validate, not a demonstrated result.
For live Parakeet, do not confuse partial captions with its separate fast-results option. The release limitations warn that fast-results mode adds repeated filler to natural speech.
Before committing, compare representative recordings on the actual target device. Check transcript errors, speaker mistakes, response delay and memory separately. Plan model installation before recording and let users replay and correct meeting transcripts. Local recognition also leaves the app’s storage and syncing choices to its developer; it is not an end-to-end privacy guarantee.
For multilingual recordings outside an Apple-local app, our guide to NVIDIA’s Nemotron 3.5 Transcribe covers a different deployment option—not another backend in this package.