New Apple Tool Lets Apps Transcribe Speech Offline in 25 Languages

A new software tool called AnySpeechRecognizer is making it easier for app developers to add speech recognition to Apple devices. Instead of being locked into Apple's built-in speech system, developers can now choose from open models that run directly on the device. This means apps could transcribe speech without sending audio to the cloud, which saves money and keeps conversations private.

The tool supports several models, including NVIDIA's Parakeet, which recognizes 25 European languages and can automatically detect which language is being spoken. Another model, Qwen3-ASR, handles 30 languages and 22 Chinese dialects, including Mandarin and Cantonese. The models are converted to Apple's Core ML format and can be downloaded from the Hugging Face Hub. They range in size from about 600 MB to 2.4 GB, so users need to download them once before use.

One standout feature is speaker diarization—the ability to tell who is speaking. Apple's built-in speech framework doesn't offer this, but AnySpeechRecognizer adds it using NVIDIA's Streaming Sortformer. This could be useful for transcribing meetings or interviews, where you want to know which person said what. The diarizer can track up to four speakers, even when they talk over each other.

The catch is that these models may not be as fast or accurate as Apple's built-in speech recognition, especially for live transcription. Parakeet is faster and suitable for live input, while Qwen3-ASR is slower and better for transcribing recorded files. Also, the models take up significant storage space, so users need to decide if the trade-off is worth it. Overall, this gives developers more flexibility and could lead to more private, cost-effective speech-enabled apps.