Apple's on-device speech recognition API, SpeechAnalyzer, surpasses Whisper Small in English benchmarks.

The development team behind the on-device AI app 'Inscribe' has compared Apple's 'SpeechAnalyzer' API, introduced for iOS 26 and macOS Tahoe 26, with OpenAI's 'Whisper' speech recognition models (Tiny, Base, and Small), and published their results.

Apple's New Speech API vs Whisper: The First Real Benchmark
https://get-inscribe.com/blog/apple-speech-api-benchmark.html

When recording meetings to create minutes, converting lecture audio into searchable text, or reviewing voice memos later, a system is needed to accurately convert long audio sessions into text. However, sending audio to an external server for processing can cause transcription to stop depending on the communication environment, and there are also concerns about sending work-related conversations or personal recordings outside of the device.

On-device speech recognition, which processes audio solely within the device, allows for transcription without relying on an internet connection. Apple has long offered the 'SFSpeechRecognizer' speech recognition API, but it was primarily designed for short audio inputs and had limitations in recognizing longer audio, such as in meetings or lectures, or when the speaker is far from the microphone.

At its Worldwide Developers Conference (WWDC25) in June 2025, Apple announced SpeechAnalyzer , the next generation of its speech recognition API. SpeechAnalyzer will be available for Apple's various platforms, including iOS 26, and will utilize a new recognition model that supports long recordings, conversations, and audio recorded from remote locations. According to Apple, the same technology is also used in features such as Notes, Voice Memos, and Journal.

Summary of Apple's annual developer conference 'WWDC25' keynote address, featuring a wealth of new information including 'iOS 26' and the new 'Liquid Glass' design - GIGAZINE

However, Apple's official announcement did not include specific figures showing how accurate SpeechAnalyzer is compared to conventional methods or other companies' speech recognition models. This left developers unable to determine whether it was worth migrating from SFSpeechRecognizer or if it would be better to use Whisper in their apps. Therefore, the Inscribe development team compared the five different speech recognition engines they use in their apps under the same conditions and published the benchmark results.

For the verification, 2,620 clear reading audios and 2,939 more difficult and noisy reading audios from the 'LibriSpeech' dataset, which is used to evaluate English speech recognition, were used. The evaluation metric was 'WER,' which indicates the word error rate, and is a numerical representation of cases where a correct word is replaced with another word, a word is omitted, or a non-existent word is added. A lower WER value indicates higher transcription accuracy.

For clear speech, the WER (Word Error Rate) was 2.12% for SpeechAnalyzer, 3.74% for Whisper Small, 5.42% for Whisper Base, 7.88% for Whisper Tiny, and 9.02% for the conventional SFSpeechRecognizer. Even with more difficult speech, SpeechAnalyzer achieved a rate of 4.56%, lower than Whisper Small's 7.95% and SFSpeechRecognizer's 16.25%. Compared to SFSpeechRecognizer, SpeechAnalyzer reduced the word error rate by approximately one-quarter.

All processing was performed on a Mac equipped with an M2 Pro processor and 32GB of memory, without sending audio to an external server. SpeechAnalyzer not only achieved higher accuracy than WhisperSmall, but also processed audio at about one-third the time per second. One hour of audio could be processed by each engine in about 1 minute 30 seconds to 5 minutes, but there was some variation in processing time because development work was also running on the same Mac during the benchmark. Inscribe says they will remeasure with the Mac idle and update the article.

On the other hand, the testing was limited to English spoken audio, and the same results may not be obtained with accented speech, meetings with multiple speakers, or audio recorded from distant locations. Based on the benchmark results, Inscribe has changed the automatic selection function of its speech recognition engine to prioritize SpeechAnalyzer for supported languages and select Whisper for others.