Apple SpeechAnalyzer API Benchmark: Performance vs Whisper and SFSpeechRecognizer

Apple SpeechAnalyzer Outperforms Whisper Small in On-Device Accuracy

Apple's new SpeechAnalyzer API, introduced in iOS 26 and macOS 26, is the most accurate on-device speech engine for English transcription among the models tested. In benchmarks using the LibriSpeech corpus, SpeechAnalyzer achieved a Word Error Rate (WER) of 2.12% on clean speech and 4.56% on noisy speech, surpassing Whisper Small, which recorded 3.74% and 7.95% respectively.

Beyond accuracy, SpeechAnalyzer is approximately three times faster than Whisper Small per second of audio. On an Apple M2 Pro (32GB, macOS 26.5.1), all tested engines ran faster than real-time, with transcription speeds ranging from 12x to 40x.

Comparative Word Error Rate (WER) Results

Engine	test-clean WER	test-other WER	Model size
Apple SpeechAnalyzer (iOS/macOS 26)	2.12%	4.56%	system
Whisper Small (WhisperKit CoreML)	3.74%	7.95%	~460MB
Whisper Base	5.42%	12.51%	~140MB
Whisper Tiny	7.88%	17.04%	~40MB
Apple SFSpeechRecognizer (legacy)	9.02%	16.25%	system

Note: Lower WER indicates higher accuracy.

Migration from SFSpeechRecognizer to SpeechAnalyzer

Developers should migrate from the legacy SFSpeechRecognizer to the new SpeechAnalyzer API immediately. The transition provides a significant accuracy boost, cutting the word error rate by 3.5 to 4 times on the same audio. Specifically, WER dropped from 9.02% to 2.12% for clean speech and from 16.25% to 4.56% for noisy speech.

In addition to raw accuracy, SpeechAnalyzer produces higher-quality output with proper punctuation and casing, whereas the legacy engine's output is rougher.

SpeechAnalyzer vs Whisper: Trade-offs and Use Cases

While SpeechAnalyzer is the superior choice for English transcription on current Apple hardware, Whisper maintains two primary advantages:

1. Language Support: Whisper supports a vast array of languages, whereas SpeechTranscriber supports approximately 30 locales.
2. Platform Agnostic: Whisper can run on any platform, while SpeechAnalyzer is restricted to Apple platforms running OS 26.

For developers shipping on-device AI, using the system-provided SpeechAnalyzer reduces app bundle size because the model is integrated into the OS rather than bundled within the application.

Benchmark Methodology and Validation

To ensure the results are reproducible and unbiased, the benchmark employed several strict controls:

• LibriSpeech Corpus: The use of LibriSpeech allows for direct comparison with OpenAI's published Whisper numbers. The benchmark results for Whisper Tiny, Base, and Small are closely align with OpenAI's published data, with a small positive offset attributed to CoreML quantization and a stricter text normalizer.
• Production Code Paths: Engines were run through the same production code used in the Inscribe app, rather than a lab harness.\n* Text Normalization: All outputs passed through a normalizer that handles casing, punctuation, and digits-to-words to ensure engines are not penalized for formatting differences.
• On-Device Verification: The harness forced on-device recognition to prevent silent fallbacks to cloud servers, ensuring a fair comparison of local compute.

Community Insights and Counterpoints

Community discussion highlighted several areas where SpeechAnalyzer may still fall short or where alternative models exist:

• Model Comparison: Some users argued that that benchmarking against older Whisper models (Tiny, Base, Small) is insufficient, suggesting comparisons against Whisper Large-V3 Turbo or newer state-of-the-art models like Nvidia's Parakeet or Mistral's Voxtral.
• Language Autodetection: Users expressed frustration with the lack of language autodetection, automatic selection of language,-language autodetection, automatic selection of language, which can lead to gibberish output if the wrong locale is not selected correctly.
• Real-world Performance: Some developers reported that SpeechAnalyzer handles diverse audio (music, multiple speakers) audio (music, multiple speakers) well but can struggle with proper nouns,
• Timestamping: While SpeechAnalyzer provides word-level timestamps for English, some users noted that other multi-ry-platform tools like ElevenLabs' Scribe v2 remains preferable for multi-language timestamping.

Limitations of the Study

• Scope: The benchmark is limited to English read speech (LibriSpeech)
• Hardware: Tests were conducted on a single machine (M2 Pro, macOS 26.5.1). While accuracy is generally consistent across Apple Silicon, processing speed will vary by chip.