pevers-whisperd-nl-whispercpp-q8_0-complete-tags

pevers whisperd-nl whisper.cpp q8_0 complete tags

This repository contains a whisper.cpp ggml q8_0 conversion of pevers/whisperd-nl, preserving the original custom speaker/event tokens from the upstream tokenizer.

Artifact

Important runtime requirement

This model is not a drop-in artifact for older stock whisper.cpp builds. The upstream model adds custom tokens after the normal Whisper timestamp range: [S1] through [S10] and (laughs).

Older whisper.cpp code assumes every token after <|0.00|> is a timestamp. That breaks decoding for this model. The Android build used for this artifact patches whisper.cpp to:

• load the full Hugging Face vocabulary including added_tokens.json
• derive special token IDs from token strings
• treat only <|0.00|> through <|30.00|> as timestamp tokens
• preserve the upstream speaker/event tokens as normal decoded tokens

Attribution

• Original model: pevers/whisperd-nl
• Base model family: openai/whisper-large-v3
• Runtime project: whisper.cpp

License

This derived conversion follows the upstream pevers/whisperd-nl license: MIT.