Representation
Streaming representation, then mute as a consumer
How Tojito turned speech understanding into a reusable layer—and made profanity filtering one consumer of it, not the whole product.
This is the architectural turn that defines where Tojito is going.
For a while, the obvious MVP story was narrow: detect profanity in live desktop audio and mute it locally. That remains a real, useful capability—and it is still what the early product emphasizes. But building only a single-purpose filter would have painted the platform into a corner. Every new idea (captions, translation, richer policies) would have meant re-deriving speech understanding from scratch, or stuffing more heuristics into the mute path.
So we separated the concerns.
Streaming speech representation
Instead of treating “hear audio → decide mute” as one inseparable blob, Tojito centers on a streaming speech representation layer: a continuous, low-latency understanding of what is being said as audio arrives. Partials and more stable finals, timing, and language context live here. The job of this layer is to describe speech as it unfolds—not to apply a product policy.
That is a product decision as much as an engineering one. A reusable representation is how a local-first audio platform grows without rewriting the core for every feature.
Profanity filtering as a consumer
Profanity filtering sits downstream. It consumes stable representation output and decides when to mute. Soft muting of the audio path remains a policy concern—separate from how speech is understood.
In plain terms:
- Audio is processed locally on the desktop.
- The representation layer turns that stream into evolving speech understanding.
- Downstream policies—starting with profanity filtering—act on that understanding.
- Other capabilities can share the same representation without inheriting mute-specific logic.
This split is why Tojito is evolving beyond a single-purpose profanity filter while staying honest about what is polished today. Filtering is available now as an early product capability. Broader speech features are direction, not a claim that every roadmap idea already ships as a finished surface.
What this is not
Representation is not marketed here as a standalone transcription or translation product. Low-latency streaming understanding exists so multiple capabilities can share one foundation. Recognition quality, captions, and translation each have their own path from “technically present or under exploration” to “ready for people.”
Hardening work around this foundation focused on making the representation-plus-policy path reliable enough to build on—not on announcing new end-user features.
For where those next capabilities sit in product language, see the roadmap. For how local processing fits the privacy story, see the docs.