PakUrdu-Conversational-ASR-ONNX (in-browser build)

Quantized (4-bit MatMulNBits) ONNX export of KhiredNetworks/PakUrdu-Conversational-ASR — the same fine-tuned weights, packaged to run fully client-side with transformers.js v4 on WebGPU. Exported from the certified merged checkpoint; quantization is the only difference from the released adapter. Certified evaluation results (7.96% WER on held-out calls, r14 — supersedes r12's 8.9%) come from the GPU run described on the model card and reproduced end-to-end in the public proof notebook Kaggle proof notebook: e1-sr-proof-r14-sr133; this build exists for the browser demo.

Use

import { pipeline } from "@huggingface/transformers";

const asr = await pipeline("automatic-speech-recognition",
  "KhiredNetworks/PakUrdu-Conversational-ASR-ONNX",
  { device: "webgpu", dtype: { encoder_model: "q4", decoder_model_merged: "q4" } });

const out = await asr(float32Audio16k, { language: "ur", task: "transcribe" });

Live demo: PakUrdu-ASR-Demo — audio is processed on-device; nothing is uploaded.

Intended use

Transcribing Urdu telephone-call audio (conversational telephony speech), decoded with language=ur, task=transcribe. Out of scope: general-purpose studio/broadcast Urdu, other languages, speaker identification, or decisions about individuals. See the main model card for the full statement.

Training data & governance

Same as the main model card: fine-tuned on real SimpleRishta customer-call audio (16 kHz telephone Urdu; transcripts PII-masked before training; caller-group-safe splits). No SimpleRishta audio, transcripts, or caller-identifying data are published here — only model weights. Published results are aggregate-only.

License

Apache-2.0, with the openai/whisper-large-v3-turbo license and attribution chain applying downstream.