Allow giving the parakeet model guidance.

[Enrico2]

It would be great if we could give the model some kind of guidance, like biasing towards using actual numbers instead of spelling out numbers.
I often say things like "two hundred thirty five" and it will write that instead of 235.
This also becomes pretty frustrating when talking about percentages. It literally writes the word percent instead of the symbol %.

[BrandonWeng]

interesting! usually this is handled in post processing, e.g

https://docs.nvidia.com/nemo-framework/user-guide/25.02/nemotoolkit/nlp/text_normalization/wfst/wfst_text_normalization.html

This is probably best handled on the client side

[Audio input]
     ↓
 Text decoder 
     ↓
Raw ASR text
     ↓
 **Post-processing**
     ├─ Punctuation restoration
     ├─ Capitalization
     └─ Inverse Text Normalization (ITN)
           ↳ "two hundred thirty-five percent" → "235%"
           ↳ "twenty twenty-five" → "2025"

Voice ink guys do something similar

https://github.com/Beingpax/VoiceInk/blob/f8332b03/VoiceInk/Whisper/WhisperState.swift

[BrandonWeng]

Closing this - feel free to re-open if you want to push back!

[nischith-memfold]

Taking up the "push back" invitation, half a year late: we hit the case post-processing can't reach, and built the decode-time version of this for the Nemotron streaming path — #841, full implementation inline.

Two things client-side post-processing couldn't do for us:

• A term the greedy decode never emits can't be repaired downstream. "Quexal" decoded as "Quex el" is fixable by fuzzy matching; the same word absorbed into a nearby real word, or dropped to blank entirely (the isolated-short-word shape in nemotron-multilingual 560ms streaming: isolated short word silently decodes to blank, controlled by the preceding audio #838), leaves nothing within edit distance to correct.
• Correction without acoustic evidence over-fires — your own CTC custom-vocabulary rescorer over-fires on short keyword terms since 0.14.8 (PR #634) #702 documented that mode on the CTC path. Decode-time shallow fusion only flips a token when its boosted logit actually beats the argmax at that step, so the acoustic evidence stays in the loop.

Measured on synthesized rare-term clips (8 targets / 4 phonetic-neighbour distractors / 2 neutrals): exact-spelling recall 0/8 → 5/8 at cbw 4.5, zero false fires (a spoken "phoenix" never typed the term "Pheynix"), zero drift on neutral speech. Happy to PR — details in #841.