sarpba
commented
10 months ago ct2-transformers-converter --model openai/whisper-large --output_dir whisper-large-v3-ct2 --copy_files tokenizer.json preprocessor_config.json --quantization int8
Valószínű meg kell adni még a modell export könyvtár otvonalát, illetve a vocabulary.json-t manuálisan kell konvertálni txt-be. mondjuk itt https://onlinejsontools.com/convert-json-to-text
Csak az alap large modell megy. Nekem a v2-v3-tól betojik az addon. (látszólag elindul szépen, de ha munkára fognám mindenféle hibát dob, lassan rájövök talán mi a baj, mert elvileg 2 hónapja már kezeli a whisper a large3-at is...
Egyébként a hugging face.n találtam egy jó small modelt, eddig az a legjobb. https://huggingface.co/domebacsi/faster-whisper-small-hu
Ez utóbbi finomhangolt egyébként nálam jobban teljesít, mint a large alap modell. Szeretnék egy medium modellt finomhangolni a CV15-el de nem tudom egyelőre, hogy elég lesz-e a gépem hozzá. Megnézem a leírást amit linkeltél.
cociweb
Hello, I've found no information about the plan of the model updates. Currently the initial multilingual fast-whisper models are really useless on my language. the WER is around 35-30% and it seems the basic commands are not recognized such as the turn on/off (in my language). Unfortunately, the recognized text is far behind the minimum viable so it does not reach the intents-> will no execute any task. I've found a guide how to fine-tune the original whisper model. - but I've never tried it. As of now, mozilla is out with Common Voice V15 and I hope the next dataset version will be available soon (around at the end of 2023)....
As I know the original whisper model is more than 1 year old (V2-large will be 1 at the end of 2023). I've found a fine-tuned model in my language (it has better accuracy but still useless) it is pre-trained with Mozilla's CV V13 dataset and can be found on the www. The time between the work of the two foundation is almost a half year. but, if we check the last half year, then it is incredibly improved: So in the last 1 year the dataset has grown 6x bigger on my language :
Additionally, As NabuCasa is the owner of some super new tts voices (Hungarian piper-Berta is much more natural then piper-Imre, piper-Anna, or Mycroft-Diana, anyways). I'm wondering that maybe some additional basic intents can be generated too - or am I thinking wrong?
But I'm really hope other non-english languages are improved as well...
Is there any plan to update these models officially? or any possibility to make the models able to be customized by our own?
+1) for testing/playing purpose unfortunately the download.py always tries to download the official models/configs/vocabs at docker startup. There is no official option to keep/start my custom model - Is there any plan to bring the possibility for optional custom models? +2) Additionally, I've did not find any guide how to compress the models to be *-int8.bin can anybody make a suggestion?