Smallest AI previews a Lightning speech update for English, Spanish and Hindi

Smallest AI made a new Lightning speech preview available for three languages.

**Dek:** Teams can opt into the new checkpoint before a planned November 28 switch. The preview gives voice teams time to test code switching and existing voice presets against their own scripts.

Smallest AI published a new Lightning v3.1 Pro text-to-speech checkpoint in preview on October 29. The change covers English, Spanish with English code switching, and Hindi with English code switching. Developers can opt into the checkpoint on any of the provider’s text-to-speech endpoints by using the preview model identifier. The changelog says the checkpoint works with the existing Pro voice presets. The published schedule gives developers a window to compare versions before the affected languages move to the new stable checkpoint. [Smallest AI model changelog](https://docs.smallest.ai/changelog/models)

Which requests change, and when

The preview identifier is `lightning_v3.1_pro_10_26`. The provider plans to move the checkpoint to the stable `lightning_v3.1_pro` identifier on November 28 for the affected languages. Its notice includes requests that omit the language field. It also distinguishes the explicit automatic-language mode, which stays on its current checkpoint, along with the other 28 languages. The preview identifier remains available until December 28 under the published schedule. An application that chooses a language explicitly can therefore have a different migration path from an application that relies on the default. [Smallest AI model changelog](https://docs.smallest.ai/changelog/models)

Existing preset compatibility gives teams a practical comparison method. Use the same voice and text in both checkpoints, and change only the model selector. A useful sample might include a Hindi sentence with an English product name, a Spanish appointment confirmation that includes an English address, and an English prompt with a date and currency amount. A release that improves one of those cases could still change pacing or pronunciation elsewhere. Listeners familiar with the affected languages can judge whether the result is clear and natural in the application’s actual context.

What to capture in an acceptance check

A team operating a customer-service line can start with a short set of frequent and high-impact prompts. Record the requested language, returned audio format, selected voice and expected reading of critical numbers. Review interruptions and playback separately when the synthesizer is part of a voice agent. A clean speech clip says little about whether a caller can interrupt it or whether a fallback message plays if the request times out.

Before the stable switch, it is useful to inventory requests without explicit languages. Those are easy to miss if configuration is spread across several services or templates. The distinction between an omitted language and an explicit auto mode is one of the release’s most consequential details. It lets teams decide which callers should hear the new checkpoint first and which flows need a comparison before the service’s default behavior changes. [Smallest AI model changelog](https://docs.smallest.ai/changelog/models)