📚 Documentation

Minimal, targeted tuning for Whisper and realtime models when the baseline already works but accuracy, segmentation, or latency still need work.

📚 DocumentationGuide

Advanced Parameter Transcription

Advanced

Advanced transcription tuning screenshot

Screenshot

What This Page Solves

Advanced parameters are for targeted correction, not for first-time setup.

Use them when:

the model route is already correct, but one class of errors keeps repeating
the output is close to usable, but not stable enough yet
you need a deliberate tradeoff between accuracy, segmentation, latency, and stability

When To Read This Page

Read it when

noisy audio causes hallucinated text
names, terms, or abbreviations are unstable
segments are too short or too long
realtime latency is acceptable but segmentation still feels wrong

Do not start here when

you have not chosen between Whisper and realtime models yet
your first baseline is not stable
you are about to change many parameters at once

Recommended Tuning Order

Change one parameter at a time, or two at most.
Re-test on the same sample clip every round.
Start from the symptom, not from the parameter list.
Promote only proven settings into a reusable preset.

Whisper: Which Problems It Usually Solves Best

Hallucination and repeated text

no-speech threshold
max context
temperature

Terminology, proper nouns, abbreviations

prompt
Beam Search or Greedy
best-of or beam-size

Only part of the source should be processed

segment transcription
offset range
length limits

Realtime Models: Which Problems They Usually Solve Best

Realtime-model tuning is usually less about decoding strategy and more about segmentation behavior:

when speech should start
when a segment should end
how much padding should be added
how to avoid fragments that are too short or too long

Typical controls include:

VAD scene presets
minimum speech and silence duration
minimum and maximum segment duration
pre/post padding and merge gap
thread count

Internally, current realtime-model inference is carried by MLEngine. In public workflow terms, you only need to understand that these are the models better suited for realtime capture.

Common Mistakes And Troubleshooting

Expecting advanced params to replace model selection They cannot fix the wrong route choice.
Changing many controls in one pass You lose the ability to explain the result.
Copying file-transcription settings directly into realtime workflows Realtime workflows must respect latency and segmentation first.
Increasing model size every time you see errors Sometimes the problem is language choice, VAD behavior, or source noise, not model size.