FlowtandBriefAI · Software · ResearchMeet Flowtand

Two Swift Packages Bring Pluggable Speech Recognition and Probability-Based Decisions to Apple Platforms

Announcement

Hugging Face

Browse Developer tools topic page →

A Hugging Face blog post announces AnySpeechRecognizer and AnyDecisionModel, two Swift packages that pair familiar APIs with a choice of open models. The speech package mirrors Apple's SpeechAnalyzer surface while adding engines such as Parakeet and Qwen3-ASR plus speaker diarization; the decision package returns typed answers and probabilities from a single forward pass rather than parsed text. The publisher reports word error rates from one Mac Studio and notes it has not yet tested on iPhone.

What the packages actually do

The publisher frames language models as only part of the machine-learning picture: a support call first needs transcription, then a routing decision. AnySpeechRecognizer keeps the shape of Apple's SpeechAnalyzer API and swaps in alternative engines. Parakeet variants and Qwen3-ASR run locally on Core ML, and the author says code using the shared API may need only an import change, with the transcriber construction differing per engine. The package also adds a speaker diarizer, which Apple's framework lacks, tracking up to four speakers via NVIDIA's Streaming Sortformer. AnyDecisionModel takes a different route: it treats a question as label probabilities read from one forward pass, returning typed answers, per-option probabilities and confidence instead of parsed generated text. The author reports that the answer to one question never feeds a later question.

A scenario for a support-engineering lead

A team handling inbound tickets could trial both packages together: transcribe recorded calls locally with a Parakeet model, attach speaker turns via the diarizer, then ask AnyDecisionModel a routing question such as which department should handle the issue and a rating question on frustration. The blog's ticket example uses a Qwen model through MLX, so a team could compare local and hosted backends for the same questions. Because decisions return probabilities, the team could route only high-confidence tickets automatically and send the rest to humans. This is a proposed experiment, not a demonstrated integration.

How to test it against your own baseline

Start with a sample of real audio and your current pipeline as the baseline. Use the package's hear evaluate command or the repository's evaluation script on your own files, as the author suggests, and record word error rates per engine. For decisions, hand-label a set of tickets and compare the returned probabilities against those labels, checking calibration rather than a single accuracy figure. Failure checks should cover accented speech, overlapping speakers and inputs outside the model's languages. Keep human review on any consequential routing until numbers hold up on your hardware and data.

Trade-offs and when to hold back

Local models mean downloads of roughly 600 MB per Parakeet, up to 2.4 GB for Qwen3-ASR and about 250 MB for Sortformer, plus memory and speed costs the author describes as higher for Qwen. The reported benchmarks are from one machine and limited datasets, and meetings were hard for every engine tested. The author has not tested on iPhone. A team should not adopt this for live, latency-sensitive transcription with Qwen, or for meeting diarization accuracy claims, without its own measurements. On-device processing may simplify privacy questions, but that is an architectural choice to confirm separately.

Context and limitations

  • The post is a short excerpt and the full article was not analyzed; details beyond the quoted text are unknown.
  • Benchmarks come from one M4 Max Mac Studio running macOS 27.0.1; the publisher says to measure on your own hardware and audio.
  • The author reports testing only on macOS 27, not on iPhone, and says meeting audio exceeded 23% word error rate for every engine tested.
  • Based on bounded source text, not independent testing. Practical suggestions are editorial proposals.