Piper TTS

Piper is the engine to test when a WebGPU path is unavailable or a CPU-oriented WebAssembly workflow is preferred. Its roughly 75MB model is smaller than Kokoro but larger than Kitten, and its 22.05kHz output is fixed. That combination can be practical for English narration drafts, accessibility prototypes, and devices where GPU support is unreliable, but it is not a promise of one speed across all processors and browsers.

Run a representative paragraph and measure both generation time and audio quality on the target device. Long text is chunked, so total time also depends on sentence structure and pauses. Keep the tab active during important work, review the complete export for repeated or dropped boundaries, and use WAV when the file will be edited further.

The Piper model exposes 904 speaker identifiers from its LibriTTS configuration. OfflineTTS lists 25 selected identifiers with readable catalog names and traits to keep the browser picker manageable. They are not 25 separately trained OfflineTTS models, verified identities of named people, or winners from a controlled listening test. The underlying voice and dataset terms still apply.

Compare speakers with the same text, speed, browser, and output format. Include proper nouns and numbers, because a pleasant sample sentence does not prove pronunciation for a real script. If a speaker is used commercially, record its identifier and review the exact upstream model card and dataset terms rather than relying only on the friendly display name.