Italian · Language-specific — CTranslate2 conversion
Converted from Corviinuss/whisper-large-v3-turbo-italian-lora, published by Corviinuss. The source authors retain credit for the fine-tune; this repository packages it for faster-whisper and Parseh.
Complete float16 CTranslate2 weights and matching tokenizer/preprocessor assets. Italian and Hindi Fast adapters are merged with their pinned Whisper Turbo base where stated in PACKAGE-PROVENANCE.json. Source revisions, conversion versions and transformations are recorded there. No additional fine-tuning was performed.
Licence and attribution
Fine-tune: MIT, as declared in the preserved SOURCE-MODEL-CARD.md. Underlying OpenAI Whisper: MIT, with its original copyright and licence in LICENSE-MIT.txt. Keep all licence files and NOTICE.txt with redistributed copies. LICENCE-SOURCES.json records where the licence texts came from. These weights are not relicensed under Parseh's GPL licence.
The source repositories supplied model cards rather than separate LICENSE/NOTICE files at the pinned revisions inspected for this publication. These declarations are the basis for distribution; they are not an independent audit of training-data rights. Training datasets are not included.
Use
Download all repository files, then use faster-whisper:
from faster_whisper import WhisperModel
model = WhisperModel("/path/to/downloaded/model", device="cpu", compute_type="int8")
segments, info = model.transcribe("audio.wav", language="it", task="transcribe", word_timestamps=True, temperature=0)
for segment in segments:
print(segment.text)
In Parseh, this package is downloaded only when the user chooses Get it in Speech to text settings, after its exact repository revision is registered in the catalogue. No conversion tools are needed by users.
Validation and limitations
Package assets and source identities are verified. This is not a claim of improved recognition accuracy, reference-word timing accuracy, or complete compatibility across recordings and platforms. New catalogue entries remain fully_compatible: false pending qualification. “Fast” and “Accuracy” describe intended trade-offs, not guaranteed accuracy gains.
- Adapter must be merged before conversion; comparative conversational evaluation is limited.
- Downloads last month
- 19