listen-windy-lingua-ps-ct2

WindyWord.ai STT — Pashto Lingua (CPU INT8 (CTranslate2))

Part of the Windstorm Labs open model catalogue. Scores, licence and attribution for every model: https://windytranslate.com/models/listen-windy-lingua-ps-ct2

Transcribes Pashto speech (Indo-European > Indo-Iranian > Iranian).

Note: EXCELLENT tier when used correctly. Derived from ihanif/whisper-medium-pashto. Verified at WER 5.3% / CER 3.2% / script-match 99.2% on 50-sample FLEURS ps_af when inference uses forced_decoder_ids (passed explicitly to model.generate() via processor.get_decoder_prompt_ids(language='pashto', task='transcribe')). With the convenience language= kwarg the model can silently drop the Pashto token and hallucinate English script on ~30% of samples (53.7% WER artifact). Always force the decoder prompt for Pashto inference.

Quality

• WER: not yet measured by us. Imported from an upstream community fine-tune.

About this variant

This is the ct2-int8 deployment format of our Pashto Lingua STT model. Load it via the ct2-int8/ subfolder.

Part of the WindyWord.ai STT fleet — covering 35+ languages that commercial speech-to-text APIs underserve, with proper dialect / script disclosures where they matter.

Usage

from transformers import WhisperForConditionalGeneration, WhisperProcessor
processor = WhisperProcessor.from_pretrained("WindyWord/listen-windy-lingua-ps-ct2", subfolder="ct2-int8")
model = WhisperForConditionalGeneration.from_pretrained("WindyWord/listen-windy-lingua-ps-ct2", subfolder="ct2-int8")

Apps

The Windy Word apps are built on this model family.

---

Provenance & License

CTranslate2 INT8 conversion of WindyWord/listen-windy-lingua-ps, whose weights derive from ihanif/whisper-medium-pashto (community fine-tune by ihanif). Lineage explicit in commit history. Licensed Apache-2.0 upstream; this redistribution is released under the same licence.