Apple SpeechAnalyzer: The New King of On-Device Transcription

Apple's new SpeechAnalyzer API significantly outperforms both its predecessor and Whisper Small in accuracy and speed for on-device English transcription. While Whisper still leads in language support, SpeechAnalyzer is now the strongest option for English on iOS and macOS. Developers are encouraged to migrate to the new API to achieve a 4x reduction in word error rates compared to legacy tools.

Key Points

• Apple's new SpeechAnalyzer is 3.5x to 4x more accurate than the legacy SFSpeechRecognizer API.
• SpeechAnalyzer outperforms Whisper Small in both accuracy and speed, running roughly 3x faster on M2 Pro hardware.
• Whisper remains the preferred choice for non-English languages and cross-platform support, as Apple's engine currently supports only about 30 locales.
• The benchmark methodology was verified by reproducing OpenAI's own published Word Error Rate (WER) numbers for Whisper models.
• The developers of Inscribe have updated their app to prioritize SpeechAnalyzer for English transcription based on these results.

Sentiment

HN is cautiously positive toward the article's core claim that SpeechAnalyzer is a serious on-device improvement for Apple users, but the agreement is heavily qualified. The dominant tone is not rejection; it is a technically engaged demand for broader comparisons and more real-world evidence. Commenters generally accept that Apple has made something important, while resisting the idea that it is categorically better than current open, cloud, multilingual, or specialized transcription systems.

In Agreement

• Apple SpeechAnalyzer appears to be a major practical upgrade over the old Apple speech stack for English transcription on modern Apple devices.
• Having a strong OS-provided local model is valuable because apps can avoid bundling their own large speech models and users get a private, shared system capability.
• Streaming partial results, low latency, and efficient use of Apple hardware make SpeechAnalyzer attractive for dictation, live captions, accessibility, and everyday transcription workflows.
• The benchmark is still useful because Whisper remains a common baseline in real apps, and the author compared engines through the same production pipeline.
• Developers and users with hands-on experience reported that SpeechAnalyzer feels faster, more accurate, or easier to integrate than previous Apple options.

Opposed

• The benchmark is incomplete because it omits many newer or larger ASR models that commenters consider more relevant than the smaller Whisper family.
• Apple's language coverage, per-language downloads, and weak language autodetection limit the result for multilingual users and mixed-language dictation.
• Real-world speech is harder than the benchmark: accents, technical jargon, noisy recordings, meetings, timestamps, speaker diarization, and poor audio may change the ranking.
• A proprietary Apple-only model is not a replacement for open or cross-platform models in self-hosted, Linux, Windows, Home Assistant, or inspectable privacy-sensitive setups.
• Raw transcription quality is only part of the product; third-party tools may still win with formatting, dictionaries, workflow polish, context awareness, history, shortcuts, and post-processing.