CoolFace
Modelpublic

taabu75/saybold-roman-urdu-turbo-ggml

sourceHugging Facemitupdated 9d agoView on Hugging Face
0likes
Model Card

Roman Urdu Turbo — GGML F16 (Preview)

On-device Urdu speech recognition with Roman Urdu (Latin-script) output. A GGML conversion of the linked MIT-licensed fine-tuning adapter, merged into OpenAI Whisper large-v3-turbo. This is an unofficial conversion, not a newly trained model.

File and compatibility

  • —ggml-roman-urdu-turbo-f16.bin: 1,624,555,275 bytes (about 1.51 GiB).
  • —SHA-256: cc50160864793d1fccd08c4c7a4ac1976588ef4f5c325959f723d78232b0d291.
  • —Whisper GGML F16; 128 mel bins, 32 encoder layers, 4 decoder layers.
  • —Use language ur, task transcribe, with translation disabled. No custom engine or prompt required.
  • —Requires a Whisper runtime that supports large-v3-turbo. Runtime working memory is larger than download size; allow approximately 2.5 GiB for this configuration.

Conversion

The base is ggerganov/whisper.cpp commit 5359861c739e955e79d9a303bcbc70fb988958b1, ggml-large-v3-turbo.bin, SHA-256 1fc70f774d38eb169993ac391eea357ef47c88757ef72ee5943879b7e8e2bc69. The adapter is Farhan-SE/whisper-large-v3-turbo-roman-urdu commit 13289b61af57158096cdbfd3817347a247696012. Eighty attention projection weights were updated using W + (64/32) × B × A, then stored in F16. Tokenizer and architecture remain unchanged. The reproducible merge script and pinned adapter configuration are included.

Verification and limitations

On 18 September 2026, an isolated runner using Saybold build 23's unmodified native speech library loaded this model and transcribed a 9.42-second public Urdu clip from Google FLEURS (ur_pk, validation row 0, ID 1537). The result used Latin letters and retained the sentence's Urdu wording. It contained recognition errors. This was one runtime compatibility check, not an accuracy benchmark or proof of reliable output across speakers, accents, noise, code-switching or longer recordings. Roman Urdu spelling varies. Review all output; script consistency and recognition quality are not guaranteed.

Training-data details and comprehensive evaluation were not supplied by the upstream adapter author. See UPSTREAM-MODEL-CARD.md for the original model card and NOTICE.txt for attribution. Saybold access plans govern its app features, not access to these MIT-licensed weights.