Hi everyone,
The project has grown into something larger than any of us perhaps imagined when we started — more hands, more moving parts, more weight behind every decision. With that in mind, I'd like to share a question that's been quietly turning over in my head, and gather everyone's thoughts before we settle on a direction.
As I've been sitting with this, three possible paths have come into focus. Each carries its own weight, its own promise, its own cost.
Path 1 — The path of restraint
We leave our current commitment to user privacy untouched, and pour our energy into craft instead: a carefully built post-processing layer — custom dictionaries, hand-tuned rules — applied before the third-party processing step. No data ever leaves the user's hands. It's the path most respectful of the trust people have placed in us. The honest limitation is that hand-built systems can only go so far; we'd be accepting a ceiling, even if a principled one.
Path 2 — The middle path
We train a dedicated post-processing model of our own, and offer it back to the community as open source — both code and weights. To make this real, we'd need to gather some user data, carefully and respectfully, to give the model something true to learn from.
Path 3 — The ambitious path
We attempt something more daring: a single end-to-end model that brings ASR and the LLM together as one. I want to be candid here — this is territory I haven't walked before, and whichever training direction we choose (Path 2 or Path 3), the model would necessarily stand on the shoulders of an open-source foundation rather than rise from nothing. Path 3 is the longer journey, with more uncertainty and more places to stumble. It, too, would ask for user data.
There's something I feel I owe everyone in plain terms: for Paths 2 and 3, data isn't a nice-to-have — it's the soil the model grows in. Without enough of it, we can't train something that truly works, and we can't even tell where the cracks are when it falters. Only Path 1 sidesteps this entirely.
If we do choose Path 2 or 3, I'd want privacy treated not as a checkbox but as a posture:
- A clear, prominent notice on the first major version that introduces data collection
- A plain-language account of what's collected, what it's used for, and what risks come with it
- Explicit opt-in — no quiet defaults — with a graceful path to decline that costs the user nothing
So the question I'd genuinely love your reflections on: of these three paths, which one feels right for us — for what we're building, and for the people who entrust us with their words?
I hold no real attachment to any one direction. I only hope we choose well, and choose it together.
Hi everyone,
The project has grown into something larger than any of us perhaps imagined when we started — more hands, more moving parts, more weight behind every decision. With that in mind, I'd like to share a question that's been quietly turning over in my head, and gather everyone's thoughts before we settle on a direction.
As I've been sitting with this, three possible paths have come into focus. Each carries its own weight, its own promise, its own cost.
Path 1 — The path of restraint
We leave our current commitment to user privacy untouched, and pour our energy into craft instead: a carefully built post-processing layer — custom dictionaries, hand-tuned rules — applied before the third-party processing step. No data ever leaves the user's hands. It's the path most respectful of the trust people have placed in us. The honest limitation is that hand-built systems can only go so far; we'd be accepting a ceiling, even if a principled one.
Path 2 — The middle path
We train a dedicated post-processing model of our own, and offer it back to the community as open source — both code and weights. To make this real, we'd need to gather some user data, carefully and respectfully, to give the model something true to learn from.
Path 3 — The ambitious path
We attempt something more daring: a single end-to-end model that brings ASR and the LLM together as one. I want to be candid here — this is territory I haven't walked before, and whichever training direction we choose (Path 2 or Path 3), the model would necessarily stand on the shoulders of an open-source foundation rather than rise from nothing. Path 3 is the longer journey, with more uncertainty and more places to stumble. It, too, would ask for user data.
There's something I feel I owe everyone in plain terms: for Paths 2 and 3, data isn't a nice-to-have — it's the soil the model grows in. Without enough of it, we can't train something that truly works, and we can't even tell where the cracks are when it falters. Only Path 1 sidesteps this entirely.
If we do choose Path 2 or 3, I'd want privacy treated not as a checkbox but as a posture:
So the question I'd genuinely love your reflections on: of these three paths, which one feels right for us — for what we're building, and for the people who entrust us with their words?
I hold no real attachment to any one direction. I only hope we choose well, and choose it together.