Work

HarshiVoice: A Sub-2M-Parameter, From-Scratch Model for Six-Language Identification and Closed-Grammar Speech Recognition on CPU, with an Analysis of Corpus-Identity Shortcuts

Year: 2026
Type: preprint
Abstract: Non-peer-reviewed preprint. HarshiVoice is a 1,973,892-parameter speech model trained from random initialisation, on CPU only, for Kannada, English, Telugu, Tamil, Hindi and Marathi. A convolution-attention encoder feeds a CTC head, which is used only for exact closed-grammar scoring of 172 command phrases (it is not an open-vocabulary recogniser; greedy CER is about 0.60), and an attentive-statistics LID head with UNKNOWN rejection. It is paired with HarshiVAD, a 56,513-parameter causal CNN voice activity detector. On held-out FLEURS test speakers (3,981 utterances), v1.3 reaches 74.8% mean per-language LID accuracy and a macro-F1 of 0.726, ranging from 98.3% (English) to 29.6% (Marathi). The int8 ONNX model (2.09 MB) matches fp32 accuracy but is 2-4x slower than fp32 on an AVX-512 CPU. The full pipeline takes a median of 63.5 ms after speech end, with 89 MB peak RSS. A speaker-disjoint probe shows that the encoder identifies the corpus of origin (FLEURS vs Kathbath) at 85-99% accuracy, and balanced sampling, gradient reversal, channel augmentation and loss re-weighting did not remove this shortcut. Every number is labelled measured-now, historical-report or not-run. Real-microphone, real-phrase and native-speaker evaluations were not run, and the project's release verdict is NO-GO. No model weights or datasets are distributed with this record.
Language: English
Open Access status: green
OpenAlex ID: W7229311952
Cites:
Cited by:
Related to: 10