---
license: mit
language: [ru]
base_model: ai-forever/sage-fredt5-distilled-95m
tags: [onnx, fp32, spelling-correction, russian]
---

# voice-sage95m-fp32

FP32 ONNX export of ai-forever/sage-fredt5-distilled-95m (SAGE Russian spelling
correction) for on-device transcript polishing in the Diktofon app.

Both the int8 and fp16 exports crash ONNX Runtime natively at inference on some
ARM phone CPUs (the low-precision gemm kernels use optional instructions).
FP32 runs on the baseline NEON sgemm kernel, mandatory on every ARMv8 core.

Files (encoder + no-KV decoder): encoder_model.onnx (~179 MB),
decoder_model.onnx (~307 MB), vocab.json, merges.txt. License MIT.