Apple's built-in SpeechAnalyzer API in iOS 26 and macOS 26 outperforms OpenAI's Whisper Small in English speech recognition accuracy while running approximately three times faster on Apple Silicon. Benchmarks published by private workspace developer Inscribe show Apple's new speech engine achieved a 2.12% word error rate on clean audio, compared to 3.74% for Whisper Small. The findings demonstrate a major leap for developers seeking high-performance, private speech recognition on Apple devices.
Benchmark Results Show Major Accuracy Gains
To evaluate speech engine performance, Inscribe ran five speech recognition models through identical production code paths on an Apple M2 Pro running macOS 26.5.1. Testing was performed using the standard LibriSpeech dataset, which includes both clean speech (test-clean) and noisier speech (test-other).
Apple's SpeechAnalyzer cut word error rates by nearly four times compared to Apple's legacy SFSpeechRecognizer API while beating Whisper Small across all tested audio samples.
The evaluation recorded the following Word Error Rates (lower percentages indicate higher accuracy):
- Apple SpeechAnalyzer (iOS/macOS 26): 2.12% clean / 4.56% noisy
- Whisper Small (WhisperKit CoreML, ~460MB): 3.74% clean / 7.95% noisy
- Whisper Base (~140MB): 5.42% clean / 12.51% noisy
- Whisper Tiny (~40MB): 7.88% clean / 17.04% noisy
- Apple SFSpeechRecognizer (Legacy): 9.02% clean / 16.25% noisy
Speed and On-Device Efficiency
In addition to delivering superior accuracy, SpeechAnalyzer proved significantly faster than OpenAI's models on Apple hardware. On the M2 Pro test system, SpeechAnalyzer completed transcription roughly three times faster per second of audio than Whisper Small.
Because the API runs entirely on-device, processing occurs locally without uploading user audio to cloud servers. Inscribe noted that the performance improvement led them to retire Whisper models in favor of SpeechAnalyzer for supported languages in their own product default settings.
Honest Caveats and Language Limits
While SpeechAnalyzer leads on English accuracy and speed on Apple Silicon, it carries notable functional trade-offs. The testing relies on the LibriSpeech benchmark, which consists of clean, read audiobook speech rather than noisy multi-speaker meeting environments.
Furthermore, Apple's SpeechTranscriber framework currently supports approximately 30 locales. OpenAI's Whisper retains a major advantage for multi-lingual applications, as it covers over 100 languages and can run cross-platform outside the Apple ecosystem.