mobilebytesensei/betterflow-whisper-hindi-medium-q4_0

betterflow — whisper-hindi-medium q4_0 (ggml)

On-device Hindi ASR model for the betterflow keyboard (privacy-first, India-first, on-device dictation). This is vasista22/whisper-hindi-medium (a Whisper-medium fine-tune for Hindi, Apache-2.0) quantized to q4_0 ggml via whisper.cpp for on-device inference. Single self-contained file, ~424 MB.

Validation (betterflow GATE-0 Phase B)

Keeps the Indic accuracy wedge at a phone-viable size — FLEURS-Hindi 8.2% WER, Common Voice-Hindi 21.7% WER, vs generic whisper-large-v3-turbo 28.5% / 39.7% on the same clips.

Usage

Load with whisper.cpp: whisper_init_from_file(...).

• Base: vasista22/whisper-hindi-medium (fine-tune of openai/whisper-medium)
• Quantization: whisper.cpp q4_0
• Runtime: whisper.cpp (ggml)
• sha256: 695119b8bae27c3ec341ff01c8c8e2f5071766aa4ec4803fe7076aa49cddd6d5