Speech recognition writes down what you said. That's rarely what you meant. "So basically I wanted to, no wait, can we meet Thursday" should become "Can we meet Thursday?"
Rules first, model second
Freshflo runs a set of deterministic fixes before and after the language model. They handle the patterns that come up again and again in spoken and Indian English, so the model has less to guess.
Measuring it
Every change is scored against a fixed test set of 76 real spoken sentences. On the day 0.10.0 shipped:
- Cloud cleanup: 76 / 76
- Apple's on-device model: 75 / 76
- Offline rules only: 57 / 76
The offline number is the one I'm pushing on.
What broke along the way
The free tier of the cloud model limits how many tokens it takes per minute, so the test run has to pace itself. When the main model is rate-limited, cleanup now falls back to a second and then a third model instead of failing.