Instructions to use iqbalaesthetic/Basira with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers.js
How to use iqbalaesthetic/Basira with Transformers.js:
// npm i @huggingface/transformers import { pipeline } from '@huggingface/transformers'; // Allocate pipeline const pipe = await pipeline('automatic-speech-recognition', 'iqbalaesthetic/Basira');
Commit History
-
remove write-access probe
Verified signature314f3d7 -
encoder(fp16): symbolic time axis — see the q4 commit for rationale and numbers
Verified signature1ac4d07 -
encoder(int8/wasm): symbolic time axis — see the q4 commit for rationale and numbers
Verified signature7aafaae -
encoder(q4): symbolic time axis so the mel window can match the buffer
Verified signatureWas [batch_size, 80, 3000], so a 6s buffer had to pad to 30s and spend ~80% of every forward encoding silence. Graph edit only, weights untouched, via scripts/make-encoder-dynamic.py: the (1500, 512) positional embedding is sliced to the activation's real length, which is what whisper does in PyTorch. Bit-identical at 3000 frames; 8.9x faster at 600 (435 -> 49 ms); teacher-forced discrimination improves 11.5x -> 16.9x.0de7ed6 -
probe: verify write access
Verified signature0a5dd25
-
ONNX exports for on-device Live Tartil: fp16 + q4 (WebGPU) + int8 (WASM)
Verified signature6fcbbaa
-
initial commit
Verified signaturef068ca1