betterflow — Whisper Hindi Small (ggml q4_0)
On-device Hindi ASR for the betterflow keyboard: vasista22/whisper-hindi-small converted to ggml + quantized to q4_0 (139 MB) for whisper.cpp.
Whisper medium q4_0 proved latency-unusable for interactive dictation on a mid-range SoC; small q4_0 with an ARM dotprod build and utterance-sized audio_ctx runs sub-real-time. WER 15.2% on FLEURS-Hindi (medium 11.1%).
- File:
ggml-small-hi-q4_0.bin· Size: 139 MB · Quant: q4_0 · Runtime: whisper.cpp arm64+dotprod - License: Apache-2.0 (from the base fine-tune)
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for mobilebytesensei/betterflow-whisper-hindi-small-q4_0
Base model
vasista22/whisper-hindi-small