| license: apache-2.0 | |
| language: [hi, en] | |
| tags: [automatic-speech-recognition, zipformer, ctc, streaming, hindi, hinglish, sherpa-onnx] | |
| # betterflow — Hindi/Hinglish streaming CTC (sherpa-onnx) | |
| Streaming zipformer2 CTC exported to ONNX for on-device dictation. Emits partials DURING speech, so | |
| first-partial latency is constant (~155 ms class) rather than proportional to utterance length — | |
| the property batch whisper structurally lacks. | |
| - **encoder.int8.onnx** (69 MB) + **tokens.txt** (104 Devanagari codepoint tokens) | |
| - Load with sherpa-onnx | |
| - Source weights: [Vengadanathan/hindi-hinglish-zipformer-ctc-streaming](https://huggingface.co/Vengadanathan/hindi-hinglish-zipformer-ctc-streaming) Phase 3 (~337h), Apache-2.0. 25.7% WER IndicVoices-hi / 39.9% Hinglish calls. | |
| - Exported + int8-quantized + verified decoding real MUCS Hinglish via sherpa-onnx 1.13.4. Casing is a non-issue: Devanagari has no case. | |
| Mirrored so betterflow can pin an immutable revision and checksum. Weights unmodified but for int8 quantization. | |