Description
First of all, thank you for this fantastic release! The model is incredibly fast and highly accurate in most scenarios.
However, there is a systemic issue regarding German language orthography: the letter "ß" (sharp S / Eszett) is not recognized during Speech-to-Text transcription. Instead, the model defaults to the Swiss-German spelling style using "ss" (e.g., converting "Grüße" into "Grüsse").
While replacing "ß" with "ss" is standard practice in Switzerland, it is grammatically incorrect for standard High German (Hochdeutsch) used in Germany and Austria for the words listed below. For business and everyday correspondence outside of Switzerland, the consistent use of "ss" renders the transcription output practically unusable without heavy manual post-processing.
As noted on the Hugging Face model card, the upstream SentencePiece vocabulary carries only 4 pieces containing "ß" compared to 58 containing "ss". This property of the v3-family tokenizer heavily biases the model towards Swiss orthography.
Environment / Context
- Repository: handy-computer/parakeet-primeline-gguf_
- Model File: parakeet-primeline-Q8_0.gguf_
- Task: Speech-to-Text (Automatic Speech Recognition)_
Steps to Reproduce
- Transcribe any German audio containing words that require "ß" after long vowels or diphthongs using transcribe-cli and the Q8_0 model.
- Observe the generated transcription text.
Expected Behavior
The model should correctly output the "ß" character for standard German vocabulary. For example, spoken words like "mit freundlichen Grüßen" must be spelled with "ß", not "ss".
Examples of Consistently Misspelled Words (Outputting "ss" instead of "ß"):
- Verbs: anschließen, fließen, gießen, grüßen, genießen, heißen, ließ, reißen, schließen, stoßen
- Nouns: Ausmaß, Fuß, Gefäß, Gruß, Grüße, Maßnahme, Maßstab, Soße, Spaß, Straße, Süßigkeit, Schließung
- Adjectives/Adverbs: außen, außer, außerdem, ausschließlich, bloß, dermaßen, draußen, dreißig, einschließlich, fleißig, groß, maßgeblich, mäßig, mutmaßlich, regelmäßig, schließlich, süß, weiß, äußerlich
Possible Solutions / Questions
Would it be possible to adjust the model or hot-fix the transcription pipeline? Since the model card mentions that the architecture supports Shallow Fusion with KenLM-based N-gram models for domain adaptation, could a dedicated language model correction layer or a simple post-processing vocabulary-mapping dictionary be implemented in transcribe.cpp to map these specific High German words back to their correct "ß" spelling?
Description
First of all, thank you for this fantastic release! The model is incredibly fast and highly accurate in most scenarios.
However, there is a systemic issue regarding German language orthography: the letter "ß" (sharp S / Eszett) is not recognized during Speech-to-Text transcription. Instead, the model defaults to the Swiss-German spelling style using "ss" (e.g., converting "Grüße" into "Grüsse").
While replacing "ß" with "ss" is standard practice in Switzerland, it is grammatically incorrect for standard High German (Hochdeutsch) used in Germany and Austria for the words listed below. For business and everyday correspondence outside of Switzerland, the consistent use of "ss" renders the transcription output practically unusable without heavy manual post-processing.
As noted on the Hugging Face model card, the upstream SentencePiece vocabulary carries only 4 pieces containing "ß" compared to 58 containing "ss". This property of the v3-family tokenizer heavily biases the model towards Swiss orthography.
Environment / Context
Steps to Reproduce
Expected Behavior
The model should correctly output the "ß" character for standard German vocabulary. For example, spoken words like "mit freundlichen Grüßen" must be spelled with "ß", not "ss".
Examples of Consistently Misspelled Words (Outputting "ss" instead of "ß"):
Possible Solutions / Questions
Would it be possible to adjust the model or hot-fix the transcription pipeline? Since the model card mentions that the architecture supports Shallow Fusion with KenLM-based N-gram models for domain adaptation, could a dedicated language model correction layer or a simple post-processing vocabulary-mapping dictionary be implemented in transcribe.cpp to map these specific High German words back to their correct "ß" spelling?