AnySpeechRecognizerTranslation site

Website Description: AnySpeechRecognizer is a native speech recognition package for Swift developers. It connects to backends such as Apple, Parakeet, and Qwen3-ASR using an interface similar to the Apple Speech API, and provides speaker separation capabilities.

What is it?

AnySpeechRecognizer is not a web service that can be used simply by uploading a recording; rather, it is a Swift package designed for developers to integrate into their applications. Its primary value lies in minimizing the need to rewrite APIs when switching recognition engines, allowing existing Apple speech processing workflows to explore different models without having to build each audio pipeline from scratch.Official Projects

This project was published by Mattt. On October 29, 2026, the author introduced both AnySpeechRecognizer and AnyDecisionModel on the official Hugging Face blog: the former handles speech-to-text conversion, while the latter is designed for structured decision-making with probabilistic outcomes. Both follow the familiar development approach of using a Swift API combined with optional models, but they are not the same product.Release Overview

What problems can it solve?

• Select an engine from the available options: By default, the Apple backend is enabled, but you can also integrate Parakeet or Qwen3-ASR, which run on Core ML.
• Processing Files and Audio Streams: Provides a unified development workflow for audio input, resource installation, and recognition results; actual output capabilities depend on the backend.
• Identifying Who Is Speaking: Use Sortformer to enhance speaker separation, making it ideal for development scenarios involving conversation logging and meeting transcription.
• Explicitly Managing Model Resources: The model is downloaded and cached by the app; developers can incorporate download prompts, storage checks, and first-run wait times into their product design.Features and Examples

Procedure for use

1. Require Swift 6.2 or later, as well as macOS 26 / iOS 26 or later; when using Parakeet or Sortformer, Apple silicon requirements must also be met.
2. Add the official repository as a dependency in Swift Package Manager, and select the version that's appropriate for your project.
3. import (data) AnySpeechRecognizer, create a recognition module and explicitly install the required resources.
4. Provide audio files or audio buffer streams to the analyzer, and update the subtitles or transcription interface based on asynchronous results.
5. Conduct an evaluation using your own language, noise environment, and equipment, and then decide on the backend and output format. The package itself does not handle audio recording; the application must handle data collection and user permissions.Installation and User Guide

Suitable Scenarios

For teams already using the Apple Speech API, this offers a path to explore alternative recognition models; for developers building meeting note-taking, recording archiving, or multilingual applications, it provides composable building blocks. It does not automatically provide complete note-editing, summarization, team accounts, or cloud synchronization features—developers still need to implement these application-level functionalities.

Comparison with Related Approaches

Reference Materials:Apple Official Documentation,Argmax Official Project. The performance of different solutions should be compared on the same device and with the same set of audio files; one should not simply combine the results presented for each solution to create a ranking.

Fees, Licenses, and Usage Restrictions

The software package uses Apache 2.0...It is not a SaaS service that charges by transcription minute. The application is still responsible for device computing power, storage, and model download costs; model weights are subject to their own licenses, and the fact that a software package is open-source does not mean that all models are licensed under exactly the same terms.Project Permit

The choice of backend affects the user experience. For example, the segmented final results from Qwen3-ASR are not the same as real-time word-by-word captions; it does not provide the same intermediate results or word-by-word timing information as Parakeet. The measurements provided by the author at the time of publication were taken in a Mac environment and have not yet been verified on an iPhone; therefore, the platform statement should not be interpreted as implying that the app has been thoroughly tested on all devices. When recording audio, conducting meetings, or handling customer calls, the app must explicitly obtain the necessary consent for recording and processing.Back-End Differences and Test Scope

Official Website: AnySpeechRecognizer Project