Apple's new SpeechAnalyzer API, benchmarked against Whisper and its predecessor
567 points
• 4 days ago
• Article
Link
Apple 在 iOS 和 macOS 26 中引入了新的语音识别 API SpeechAnalyzer，实际上取代了旧有的 SFSpeechRecognizer 。由于 Apple 没有公布此次更新的官方准确性基准，开发者和用户对于其相较于 OpenAI 的 Whisper 等已有方案的性能存在疑问。为此，研究者用来自 LibriSpeech 的 5,559 条测试语句，比较了新的 Apple 引擎、旧版 API 以及三种完全在设备端运行的 Whisper 变体，开展了一次全面的基准测试。
结果显示，SpeechAnalyzer 目前是 Apple 平台上最准确的本地语音引擎。它在所有测试的 Whisper 模型（包括 Whisper Small）之上，并且效率显著更高，运行速度约为其三倍。相比之下，旧版 SFSpeechRecognizer 表现最差，其准确度甚至低于最小的 Whisper 模型。对于那些不只依赖简单语音命令的开发者，数据表明迁移到 SpeechAnalyzer 是非常值得的，因为它在转录准确性上有显著提升，且输出文本更干净、带有标点。
在与 Whisper 的对比中，Apple 的新引擎在现代 Apple 硬件上成为英语转录的更优选择。虽然 Whisper 在语言覆盖面上仍有优势，支持 100 多种语言，而 Apple 的 SpeechTranscriber 约支持 30 种，但在英语处理方面两者的差距已大幅缩小。因此，在英语任务中，Whisper 不再是追求最高本地准确度时的默认选项。基准测试方法严谨，采用了标准的 LibriSpeech 语料，并确保 Whisper 的测试结果与 OpenAI 公布的指标一致，从而验证了针对 Apple 引擎所用方法的有效性。
该基准的透明性体现在原始转录的公开以及使用相同的生产代码路径，确保结果反映真实使用场景而非孤立的实验室条件。研究还指出，所有被测引擎的运行速度均远快于实时，但新的 Apple API 在效率上具有明显优势。这项分析已推动实际的软件部署调整，例如促使 Inscribe 应用在支持的语言中优先采用更准确的 SpeechAnalyzer 。
尽管收益显著，结果也有局限：LibriSpeech 主要以英语朗读语音为主，后续需要测试带口音的语音、远场录音或多说话人的会议音频等场景，以评估各引擎的表现。此外，虽然预计这些准确性结论在 Apple Silicon 系列硬件上普遍适用，但运行速度会随具体芯片架构有所波动。总之，对于当前的 iPhone 和 Mac 用户来说，系统内建工具现在能在不牺牲准确性的前提下，提供一流且注重隐私的本地语音转录解决方案。
Apple has introduced a new speech recognition API, SpeechAnalyzer, as part of iOS and macOS 26, effectively replacing the older SFSpeechRecognizer. Because Apple did not provide official accuracy benchmarks for this update, there has been significant uncertainty for developers and users regarding its performance compared to existing solutions like OpenAI's Whisper models. To address this, a comprehensive benchmark was conducted using 5,559 test utterances from the LibriSpeech dataset, comparing the new Apple engine, the legacy API, and three Whisper variants, all running fully on-device.
The results clearly indicate that the new SpeechAnalyzer is the most accurate on-device speech engine currently available on Apple platforms. It outperformed every Whisper model tested, including Whisper Small, while also demonstrating significantly higher efficiency, running approximately three times faster. In contrast, the legacy SFSpeechRecognizer performed the poorest, proving to be substantially less accurate than even the smallest Whisper model. For developers relying on the old API for anything beyond simple voice commands, the data suggests that migrating to SpeechAnalyzer is highly recommended due to the significant gains in transcription accuracy and the output of cleaner, punctuated text.
When evaluated against Whisper, Apple's new engine emerged as the superior choice for English transcription on modern Apple hardware. While Whisper maintains distinct advantages regarding language support, with compatibility for over 100 languages compared to the roughly 30 supported by Apple's SpeechTranscriber, the gap for English processing has closed. Consequently, for English-language tasks, Whisper is no longer the automatic default for those seeking the highest accuracy on-device. The benchmarking process was robust, utilizing standard LibriSpeech corpora and ensuring that the Whisper results aligned with OpenAI's own published metrics, which serves to validate the methodology used for the Apple engine tests.
The transparency of this benchmark is reinforced by the publication of raw transcripts and the use of identical production code paths, ensuring that the findings reflect real-world usage rather than isolated laboratory conditions. The study highlights that all tested engines operated comfortably faster than real time, though the efficiency of the new Apple API provides a distinct performance edge. This analysis has already led to practical changes in software deployment, specifically prompting updates to the Inscribe application to prioritize the more accurate SpeechAnalyzer for supported languages.
Despite the clear benefits, there are noted limitations to these findings, as the LibriSpeech dataset focuses primarily on English read speech. Future testing will be necessary to determine how these engines perform with accented, far-field, or multi-speaker meeting audio. Furthermore, while these accuracy results are expected to hold across Apple Silicon hardware, performance speeds will naturally fluctuate based on the specific chip architecture. Ultimately, for users on current iPhones or Macs, the built-in system tools now provide a best-in-class, privacy-conscious solution for local speech transcription, removing the need to compromise on accuracy for the sake of on-device processing.
238 comments • Comments Link
- 最先进的 speech-to-text 技术已远超基础版 Whisper；Nvidia 的 Nemotron 、 Parakeet 以及 MOSS-Transcribe-Diarize 等较新方案在嘈杂或多语环境下表现更佳。
- 转录模型的适用性高度依赖具体场景：有些模型注重逐字还原和碎片信息的精确性，另一些则更强调输出的流畅性和可读性。
- Apple 原生的设备端转录（通过 SpeechAnalyzer API）相比传统系统有显著进步：借助 Neural Engine 等专用硬件，字错率更低，能效更高。
- 关于 Apple 原生工具与专业或专用第三方替代方案的可用性，社区仍有争议，尤其是在术语处理、格式化以及跨方言长期保持准确性方面。
- 许多开发者和高级用户认为，Apple 原生工具虽然便捷，但在 diarization 、多语言检测和自定义术语词典等方面缺乏精细控制。
- "vibe-coded" 的 Whisper 封装普及导致市场上出现大量低质应用，这些产品常常忽视人机交互准则，促使用户转向像 Handy 或 Wispr Flow 这样更成熟、功能更丰富的 FOSS 替代方案。
- 地区口音（包括 UK 、 Australia 和 New York 的口音）仍是主流 STT 模型的长期挑战，经常迫使用户调整说话方式才能获得可接受的识别准确性。
- 有人建议将 Apple 的专有设备端模型逆向工程并移植到其他平台，但也有人认为这在技术上不可行，并低估了现代针对特定硬件高度优化的模型权重的复杂性。
- 依赖 Word Error Rate (WER) 等单一简化指标可能具有误导性：错误率降低四倍，并不必然带来日常使用体验的四倍提升。
- 把模型拿来与过时的 Whisper 版本做基准测试很普遍，但鉴于 Whisper-Large-V3-Turbo 等更快更准的更新版本以及近期的 open-weight 替代品，这种做法正越来越被认为不合适。
speech-to-text 技术的前景正迅速从通用、依赖云的模型转向在设备端利用专用芯片进行高度优化的实现。尽管 Apple 持续缩小原生 OS 功能与第三方工具之间的差距，讨论仍显示出集成功能的便捷性与高级用户及小众专业人士所需高性能之间的明显张力。新推出的专有引擎在速度和效率上表现出色，但市场仍强烈要求更高的透明度、更好的技术术语处理能力以及对多种口音和语言的稳健支持。随着封装程序门槛降低，未来很可能偏向那些通过巧妙的后处理和卓越的界面设计，成功解决可用性"最后一英里"的开发者。 • State-of-the-art speech-to-text has moved beyond basic Whisper models, with newer options like Nvidia's Nemotron, Parakeet, and MOSS-Transcribe-Diarize offering superior performance, especially in noisy or multilingual environments.
• The effectiveness of a transcription model often depends on the specific use case, as some models prioritize literal, fragment-accurate transcription, while others emphasize clean, smoothed-out output.
• Apple's native on-device transcription via the SpeechAnalyzer API shows significant improvements over legacy systems, offering lower word error rates and better power efficiency by leveraging dedicated hardware like the Neural Engine.
• Disagreement persists regarding the "usability" of Apple's native tools versus professional-grade or specialized third-party alternatives, particularly concerning technical jargon, formatting, and consistent accuracy across various dialects.
• Many developers and power users find that while Apple's native tools are convenient, they often lack granular control over features like diarization, multi-language detection, or custom terminology dictionaries.
• The ubiquity of "vibe-coded" Whisper wrappers has created a saturated market of low-quality apps that often ignore established human interface guidelines, leading users to seek out more polished, feature-rich, or FOSS alternatives like Handy or Wispr Flow.
• Regional accents, including those from the UK, Australia, and New York, remain a persistent challenge for mainstream STT models, frequently forcing users to modulate their speech patterns to achieve acceptable accuracy.
• While some enthusiasts suggest reverse-engineering Apple's proprietary on-device models to port them to other platforms, others argue this is technically impractical and ignores the complexity of modern, highly optimized, hardware-specific model weights.
• Using simplistic metrics like Word Error Rate (WER) can be misleading, as a 4x reduction in error does not always translate to a perception of "four times better" in practical, daily usage.
• Benchmarking models against outdated versions of Whisper is common but increasingly seen as a poor practice, given the availability of newer, faster, and more accurate models like Whisper-Large-V3-Turbo and recent open-weight alternatives.
The landscape of speech-to-text technology is shifting rapidly from general-purpose, cloud-dependent models toward highly optimized, on-device implementations that leverage dedicated silicon. While Apple continues to close the gap between native OS functionality and third-party tools, the discussion reflects a clear tension between the convenience of integrated features and the high performance required by power users and niche professionals. Despite the impressive speed and efficiency of new proprietary engines, there is a strong demand for more transparency, better handling of technical vocabulary, and robust support for diverse accents and languages. As the barrier to building wrappers lowers, the future of the space appears to favor developers who can successfully bridge the "last mile" of usability through clever post-processing and superior interface design.