FluentWhisper ✨

A 208-tensor LoRA adapter that teaches whisper-large-v3-turbo to skip the fillers, repetitions, and false starts that clutter raw transcripts.

Compare vanilla Whisper with the adapter side by side — the diff pane highlights every deletion in red strikethrough: discourse markers ("well", "you know"), repetitions, and self-repairs. Vanilla already drops most "um/uh"; this adapter handles the rest.

Apache-2.0 3.4% WER vs 9.4% vanilla N=250 · ±~1pp CI No API key · no cloud
⚠ Caution: the adapter deletes repeated spans aggressively — phone numbers, repeated digits, and intentional emphasis ("very, very good") can be silently dropped. See the gallery for a live example.

Examples (real test-set results)

Self-repair + discourse marker

Vanilla: Well, McNew, New and Lara, I guess is leaving. One of them is leaving.
Cleaned: well Macneau and Lerre I guess is leaving one of them is leaving

Well , McNew , New and Lara , I guess is leaving . One of them is leaving .

Drops "you know" + "I mean"

Vanilla: Or, you know, to find a job where they can learn how to support themselves. I mean, I guess we're getting kind of off the subject here, but...
Cleaned: or to find a job where they can learn how to support themselves I guess we're getting off the subject here but

Or , you know , to find a job where they can learn how to support themselves . I mean , I guess we ' re getting kind of off the subject here , but . . .

Repetition removal

Vanilla: The attitude of the staff, as you said, is really very, very important.
Cleaned: the attitude of the staff as you said is really very important

The attitude of the staff , as you said , is really very , very important .

Single "you know" removal

Vanilla: The governor, you know, has been trying to decide whether he's going to commute it or not.
Cleaned: the governor has been trying to decide whether he's going to commute it or not

The governor , you know , has been trying to decide whether he ' s going to commute it or not .

Limitation: intentional repetition is lost

Vanilla: It was very, very good.
Cleaned: it was very good

It was very , very good .

Note: long clips (>30s) are split into 30s windows and concatenated. Number sequences (e.g. phone numbers) can be mishandled by the cleaner — see the limitations note.