LM likelihood

LM likelihood suggests words by comparing their fit with the surrounding transcript. It runs locally, using a model you already installed. Whisper and the connected LLM methods have separate settings. It is an experimental correction method; check suggestions by listening.

1. Choose a model#

Open Settings → LM likelihood, press Refresh installed models, choose a model and press Use selected model. Model selection is available from any admitted browser. The selected model also appears in the review’s Models, preferences & skills panel.

The Candidate cutoff chooser offers No cutoff, 0.1 — suggested by preliminary experiments, or Custom cutoff between 0 and 1. Press Save review options to keep your choice. The cutoff affects only extra model suggestions; Whisper’s original and genuine alternatives are always compared. The suggested 0.1 is a starting point, not a proven accuracy setting.

For the first setup, expand Setup & advanced options on the Parseh computer. Choose Installed Unsloth GGUF, Installed Ollama model, or an explicit local model path. Save the settings before refreshing the list. Parseh uses existing weights read-only: install models yourself in Studio, Ollama or another application. A model on another computer must also be accessible to this local worker.

The local scoring program is installed separately with Install / rebuild isolated runtime. It needs a C/C++ compiler; GPU builds also need the matching hardware toolkit. Choose CPU with 0 GPU layers, or a supported GPU backend with −1 for all layers. Its memory use is separate from Studio or Ollama, even when they share the same model file. Parseh does not unload their models. Cancel scoring and unload worker stops Parseh’s scoring work.

Complete single-file text-generation GGUF models are supported, including Ollama backing files without a .gguf filename. Split weights and required adapters are refused. Models labeled text only use their complete text model without optional vision files.

2. Review words#

After Whisper finishes, choose LM likelihood. Click, focus or tap a word to inspect the alternatives. The original and every genuine Whisper alternative are compared alongside additional words found by the local model. No chatbot prompt or assistant response is used.

Prefer similar-sounding words starts enabled. Turn it off for a second pass on unresolved words. The filter uses dictionary readings where available and otherwise a spelling approximation. When pronunciation cannot be inferred, such as an unfamiliar Han character, it keeps the candidate available. It never removes Whisper’s original or supplied alternatives. Turning it off keeps the saved probability cutoff; lower it with the Candidate cutoff chooser if you need a wider search.

Choose Select a section, then its first and last word. Methods review only unlocked targets inside that section while keeping the surrounding context. For one word, use Review this word with LM likelihood in its inspector; this also works for a word Whisper did not flag.

Accept an alternative or enter a correction yourself. Select best for all chooses the latest method’s first choices in the pending draft, skipping locked words, incomplete comparisons and tied first choices. Listen before trusting these choices: a fluent word can still be wrong. I’m sure · lock word keeps a checked word through later methods; unlock it to reconsider.

Save & pause keeps the pending transcript, decisions, edits and locks on the Parseh computer. Reopen it with Continue pending transcription in Videos, even after a server restart. Only Use this transcript fills the transcript box. Use or discard removes the saved review. Caption timing stays unchanged; changed word timing needs review in the existing timing editor.

3. When a comparison is incomplete#

If a word cannot be fully compared, its original remains available. Other words can still finish. Retry unresolved words or edit them yourself. The ordinary inspector shows choices; Technical details and LLM responses hold the numerical scores, model identity, bounded context and failure reasons. These scores measure language-model fit, not the probability of having heard the right word.

Bounded word search can miss useful alternatives, especially in languages without spaces. Single-word comparison cannot resolve interacting mistakes jointly. The installation’s docs/lm-likelihood.md explains the numerical rule and limits.