The 131M-parameter Scylla's Band v4 can generate its first audio segment on an iPhone 17 Pro Max in roughly 70 milliseconds, with support for 7 languages and 14 voices. Our view: on-device speech generation is moving from the question of whether it can run to whether it is fast enough and genuinely deployable.
What This Is
Scylla's Band is an open-source text-to-speech (TTS) model independently developed by solo developer lowkeytea over nearly a year. Beyond standard text-to-speech, it can generate singing from sheet music, with versions for PyTorch, ONNX Runtime, WebGPU, and Core ML, plus sample code for mobile devices.
The model supports English, Spanish, French, German, Italian, and Vietnamese. Users can control speaking rate, pitch, and intensity, and have multiple voices engage in dialogue. The developers report a 0.85% English word error rate, although its leaderboard submission is still pending; singing is currently supported only in English.
Industry View
Its most valuable feature is not just model quality but its delivery path: the same capabilities cover desktop, mobile, and browser environments, with open weights and runtime code. This lets developers embed voice features in applications without relying on cloud APIs, while also improving privacy, cost control, and offline use cases.
The risks are equally clear: the project is maintained by one person, limiting its ecosystem and long-term support capacity; the voice may still distort during fast-paced passages or large pitch changes; and 14 fixed voices do not cover the customization needs common in enterprise deployments. More importantly, its test results still require independent verification, while stability, concurrency, and power consumption in real applications remain to be observed.
Impact on Regular People
For enterprise IT: Scenarios requiring continuous speech output, such as customer service, training, and accessibility, may gain a lower-cost, lower-latency local option. Before deployment, however, teams still need to assess maintenance, licensing, and performance consistency.
For individual professionals: Meeting transcription, presentation voiceovers, and first drafts of content can be produced more offline, without uploading every piece of sensitive speech to the cloud. The model's ability to sing and conduct multi-character dialogue also lowers the barrier to creating audio content.
For the consumer market: Offline translation, voice assistants, and companion applications will place greater emphasis on on-device responsiveness. Still, fixed voices and occasional mispronunciations mean that in the short term, it is better suited to serving as an embeddable foundational capability than a complete replacement for mature commercial services.