Hello, I’ve fine-tuned this version of Whisper for better Romanian transcription but when trying to deploy it I run into a problem. Usually, I load the model and the processor with WhisperForConditionalGeneration and WhisperProcessor respectively, and finally get the transcription with the model.generate() function from the transformers library (to which I would give as parameter the language, to avoid making Whisper guess it), but when using the serverless API from thf, I can’t change or specify the language so it tries to guess (and sometimes misses completely). Is there any way that would allow me to specify that I want a transcription in Romanian? If not, are there any alternatives?
Related topics
Topic | Replies | Views | Activity | |
---|---|---|---|---|
Finetuned whisper model translating instead of transcribing | 2 | 601 | December 31, 2023 | |
How to set language in Whisper pipeline for audio transcription? | 2 | 6257 | June 22, 2023 | |
How to fine-tune whisper on unsupported language? | 1 | 40 | October 12, 2024 | |
Language detection with Whisper | 15 | 19972 | August 30, 2024 | |
How to use Whisper from huggingface for ASR | 0 | 515 | June 21, 2023 |