为什么值得关注

该项目提供罕见的学者级梵语语音识别、诚实的实验报告（包含失败案例），以及利用用户同意数据将真实环境字错误率从24%降至15.6%的数据飞轮，同时保持开源并可部署。

适合谁使用

梵语学者及经典研究者
吟诵练习者与教师
从事低资源语言语音识别的研究人员
构建交互式语言工具的技术开发者

典型使用场景

高精度转写吟诵和经典散文
通过音节级反馈和格律感知分割进行吟诵练习
收集经用户同意、分级可信的数据以持续改进共享语音识别模型
使用连音归一化词错误率评估语音识别，解决梵语词边界问题

项目优势

在录音室吟诵数据上达到6%字错误率，真实环境降至约15.6%
透明记录成功与失败实验，包含详细诊断
提出连音归一化词错误率指标，揭示梵语词错误率一半源自空格
提供支持多种文字的交互式网页工具（Vāgbodhinī 和 Su-śrotā），并形成自我改进的数据飞轮

使用前须知

声学模型在干净录音室音频上出现饱和，进一步改进主要依靠任务重构（验证而非识别）
额外的人工标注数据未能降低字错误率，反而因空格漂移损害词错误率
基于规则的切分和后处理方法（分词器、ByT5）因梵语能产的连音而无效

README 快速开始

Su-śrotā — Scholar-grade Sanskrit ASR & chant practice

Sanskrit speech recognition tuned for śāstric / recitational Sanskrit, plus Vāgbodhinī, an interactive chant-practice tool built on it. This repository documents the full set of experiments — including the ones that did not work, which were as instructive as the ones that did.

ASR model: IndicConformer-CTC (Sanskrit slice). "v5" is the studio-finetuned base (on HF); v11-ep3 is the current deployed model — v5 continued on real user data (see §5b).
Live tools: Vāgbodhinī (chant/prose tutor) and Su-śrotā (dictation/transcription) — both serve the same shared core ASR and both collect consented data that improves it.
Author: Prof. Prathosh A P, Indian Institute of Science, Bengaluru.

Headline result (§5b): the model was only saturated on clean studio audio. On real, in-the-wild user audio it had huge headroom (24% CER) — and a small batch of consented user recordings closed a third of it (24% → 16% CER, 86% → 34% WER) with no loss on the studio benchmarks. The data flywheel works.

1. Motivation

Off-the-shelf Sanskrit ASR is trained on conversational IndicVoices-style data and degrades badly on recitation and śāstra — dense compounds, sandhi, retroflex/aspirate contrasts, pitch, and long metrical utterances. The goal here was a model and tooling good enough for scholars: accurate on chant and prose, and useful as a practice aid rather than a transcription toy.

Two things shaped everything:

Segmentation dominates Sanskrit WER. Word boundaries in Sanskrit are largely a writing convention (sandhi fuses words). Standard WER punishes boundary disagreements that are not errors. We therefore report a sandhi-normalized WER (SN-WER) alongside CER/WER.
The acoustic model saturates. Past a point, adding data stops moving CER; the leverage moves to how you use the model (known-reference tasks, decoding, tooling), not how you train it.

2. Data & preprocessing

Recitation + śāstra prose corpora (Bhāgavata chant, Vedānta prose), a TTS-speaker slice, and a multi-scholar annotation set collected via a purpose-built web portal.
Content-only normalization: NFC, keep Devanāgarī letters + combining marks, dro

项目描述

Scholar-grade Sanskrit ASR (finetuned IndicConformer-CTC) + Vāgbodhinī metre-aware chant-practice tool

sushrota-sanskrit-asr