Vosk

Vosk is the workhorse for embedded ASR. It will not match Whisper for a finished podcast transcript, but it will run on a Raspberry Pi without internet, on an Android phone offline, or inside a desktop application that needs voice input without sending audio to the cloud.

For developers building offline-first apps where Picovoice's commercial licensing is a blocker, Vosk is the open-source default. The platform support is the most generous in open-source ASR: bindings for Python, Node, Java, C#, Go, iOS, Android, and direct linking against C++ are all maintained.

Models range from tiny (under 50MB, suitable for embedded use) to larger ones (around 1.5GB, closer to commercial quality but still meaningfully behind Whisper-large). The toolkit is built on Kaldi, which is mature speech recognition infrastructure but pre-dates the transformer era; the accuracy gap to Whisper is real and visible on conversational or noisy audio.

The trade-off is exactly the one you'd expect. Vosk's models are small enough to run anywhere, while Whisper's large models need real compute. For applications where 'works offline on a constrained device' is the requirement, that trade is worth making.

For applications where 'highest possible accuracy' is the requirement, host Whisper instead. The Apache 2.0 licence is permissive enough for commercial use without legal review.