A new open-source Swift package gives Mac developers a practical way to add local speech transcription and speaker labels without sending recordings to a cloud service. AnySpeechRecognizer 0.1.0 shipped on October 28, and the model files and source code are available now.[3][4][7]
The important qualifier is “developers.” This isn’t a polished meeting-notes app. It’s a library and command-line tool for people building one. If that describes you, though, the package puts several useful pieces behind an API that closely resembles Apple’s own modern Speech framework.[6][7]
AnySpeechRecognizer can use Apple’s built-in transcription engine, NVIDIA’s Parakeet speech models, or Alibaba’s Qwen3-ASR models. It also adds NVIDIA’s Streaming Sortformer for speaker diarization, the part that decides who spoke when.[1][7] Everything can run on the device after the selected model files have been downloaded.[7]
Audio deserves local treatment more than most AI inputs. Meetings, interviews, medical notes and family recordings are exactly the files many people would rather not upload. Local processing keeps the recording and transcript under the user’s control, although any app built on the package still needs its own sensible storage and deletion policy.[7]
What shipped
The package mirrors Apple’s `SpeechAnalyzer`, `SpeechTranscriber`, `DictationTranscriber`, `SpeechDetector` and `AssetInventory` interfaces. In the simplest case, a developer changes the import from Apple’s `Speech` framework to `AnySpeechRecognizer`, then chooses a different model when creating the transcriber.[1][7]
Apple’s own `SpeechAnalyzer` arrived on macOS 26 and iOS 26. It accepts files or streamed audio, installs required assets and returns results asynchronously.[6] AnySpeechRecognizer keeps that programming model while opening the engine choice.
Parakeet is the best fit for fast transcription. The multilingual Parakeet TDT 0.6B v3 conversion recognizes 25 European languages, uses int8 encoder weights and float16 components, and carries a CC BY 4.0 license.[4] The package can process files or live audio, return provisional text while someone is speaking, and attach word timing and confidence.[7]
Qwen3-ASR covers a different language set. Its 0.6B and 1.7B versions recognize 30 languages and 22 Chinese dialects, including Mandarin and Cantonese, plus Japanese and Korean.[1][7] The trade is speed and interactivity: Qwen produces a final result for segments of up to 40 seconds, lacks word-level timing in this package, and runs on the Mac’s GPU rather than the Neural Engine.[7]
Speaker labeling comes from a separate 246MB Sortformer model. It can track as many as four speakers, including overlapping speech, with about 1.04 seconds of algorithmic latency in the Core ML conversion.[5][7] Five people in one room is where the neat demo starts to look more like an actual meeting.
The storage and memory bill
The package’s Parakeet v3 files take about 632MB. Asking for faster live results adds a second 591MB encoder, and speaker diarization adds about 246MB.[7] Add those documented inputs and a complete live-transcription setup is roughly 1,469MB, or about 1.4GB. File transcription with speaker labels is lighter at about 878MB.
Those are download and storage figures, not a claim that the whole 1.4GB remains resident in memory. In the developer’s M4 Max tests, Parakeet used about 100MB of peak memory after compilation; the first Core ML compilation could briefly reach about 680MB.[7] Sortformer used about 135MB on its default CPU-and-Neural-Engine path, while a faster CPU-and-GPU setting used about 530MB.[7]
The larger Qwen option is more demanding. Its 1.7B conversion is a 2.36GB download and needed about 4.4GB of memory in the author’s tests. The 0.6B version is 0.94GB and used about 1.7GB.[7] That still fits easily on current Apple-silicon Macs, but it changes the answer for an iPhone app or a Mac already juggling a large local language model.
Hardware support is narrower than the phrase “Apple platforms” suggests. The package requires macOS 26 or iOS 26, Swift 6.2 or newer, and Apple silicon for Parakeet and Sortformer.[7] The author tested the speech benchmarks on an M4 Max Mac Studio running macOS 27.0.1, not across the Mac lineup and not on a physical iPhone.[1][7]
The speed claim needs context
On that M4 Max, the developer reports that Parakeet processed one hour of recorded audio in 18 to 31 seconds in file mode. Live-mode simulation took 93 to 112 seconds per hour of audio.[7] Those figures work out to roughly 116x to 200x real time for files and 32x to 39x for the live path.
Accuracy looked strong on clean read speech. The reported word error rate for Parakeet v3 was 1.55% on LibriSpeech test-clean files lasting three to eight minutes, versus 2.12% for Apple’s transcriber in the same project test.[1][7] Meetings were much uglier: every tested engine exceeded 23% word error rate on AMI recordings with overlapping speakers.[1]
These are reproducible project benchmarks, not independent HighVRAM tests. The repository includes an evaluation command and scripts for rerunning the measurements, which is considerably better than a benchmark chart with no method attached.[7] It still doesn’t tell us how an M1 MacBook Air handles a noisy kitchen-table conversation.
Who should use it
Developers building private transcription, interview, accessibility or voice-note tools on Apple silicon now have working source code and downloadable models rather than a research demo.[3][4][7] Parakeet is the sensible first choice for English and supported European languages because it offers live partial results, word timing and much higher reported throughput. Qwen3-ASR earns its extra memory when its broader Asian-language coverage is the requirement.
For ordinary users, the advice is to wait for applications built on top of the package unless you’re comfortable with Swift and the command line. The included `hear` tool can download models, transcribe local files, add speaker labels and report evaluation metrics, but setup still involves building a Swift package and choosing explicit backends.[7]
The release makes local Mac transcription easier to build, not finished. Physical-iPhone testing, performance on base-model Macs and independent accuracy checks on messy real recordings will decide how far it travels beyond developer machines.[7]