📋 摘要
⭐ 聚焦Indic TTS工程改造，与SE for AI、公平性测试、形式化方法等兴趣关联很弱。
Praxy Voice: Voice-Prompt Recovery + BUPS for Commercial-Class Indic TTS from a Frozen Non-Indic Base at Zero Commercial-Training-Data Cost
中文
针对商业级 Indic TTS 与最佳开源基座（Chatterbox、Indic Parler-TTS、IndicF5）在音系维度上的差距，尤其是广泛采用的多语言基座 Chatterbox 甚至无法 tokenise Telugu 与 Tamil 的问题，本文探究在不训练新声学 decoder、不使用任何商业 TTS 训练数据的前提下，将非 Indic-native 基座提升至商业级 Telugu、Tamil、Hindi 输出所需的最小干预。方法包含三部分：(1) BUPS（Brahmic Unified Phoneme Space），将七种 Indic 文字确定性地罗马化为 ISO-15919，使 Chatterbox 的 Latin tokeniser 可处理；(2) 仅在文本 token 预测器 (t3) 上施加 LoRA adapter，使用约 1,220 小时已授权 Indic 音频并以 Hindi-proxy language_id 训练；(3) voice-prompt recovery 配方，即 8–11 秒同语种参考片段加三项 sampling 覆盖（Config B：exaggeration 0.7、temperature 0.6、min_p 0.1）。鉴于 LoRA 在 Hindi 上反而损害准确率，作者采用 vanilla Chatterbox + Config B，构成双分支部署。在 10 句 pilot 集与配套 PSP benchmark 上，Praxy Voice 持平或略优于商业基线：Telugu retroflex collapse 26.7%（Sarvam Bulbul 33.3%）、Tamil-zha collapse 71%（商业三家 86%）、Hindi LLM-WER 0.025（与 Cartesia Sonic-3 持平）。对句内 code-mix 增设第三分支（IndicF5 + 原生字符 transliteration），将 Hi/Te/Ta 的 code-mix LLM-WER 由 0.80–0.85 降至 0.14–0.27。作者公开 R6 LoRA 权重 (Apache-2.0)、推理代码与 router (MIT) 及 Gradio demo。
English abstract
Commercial TTS systems produce near-native Indic audio, but the best open-source bases (Chatterbox, Indic Parler-TTS, IndicF5) trail them on measured phonological dimensions, and the most widely adopted multilingual base (Chatterbox, 23 languages) does not even tokenise Telugu or Tamil. We ask: what is the minimum intervention that brings such a non-Indic-native base to commercial-class output on Telugu, Tamil, and Hindi, without training a new acoustic decoder and without any commercial TTS training data? We combine three pieces: (1) BUPS, a Brahmic Unified Phoneme Space that deterministically romanises seven Indic scripts to ISO-15919 so Chatterbox's Latin tokeniser can process them; (2) a LoRA adapter on only the text-token predictor (Chatterbox's t3), trained on ~1,220h of licensed Indic audio with a Hindi-proxy language_id; (3) a voice-prompt recovery recipe -- an 8-11s same-language reference clip plus three sampling overrides (exaggeration 0.7, temperature 0.6, min_p 0.1; "Config B") -- that recovers commercial-class acoustic output with no acoustic-decoder training. On Hindi, the LoRA regresses accuracy and we instead use vanilla Chatterbox + Config B, giving a two-branch deployment. Evaluated on 10-utterance pilot sets with the companion PSP benchmark, Praxy Voice matches or slightly leads commercial baselines: 26.7% retroflex collapse on Telugu (vs Sarvam Bulbul 33.3%), 71% Tamil-zha collapse (vs commercial trio's 86%), 0.025 LLM-WER on Hindi (tied with Cartesia Sonic-3). For intra-sentential code-mix we add a third branch (IndicF5 + native-script transliteration) that drops code-mix LLM-WER from 0.80-0.85 to 0.14-0.27 across Hi/Te/Ta. We release R6 LoRA weights (Apache-2.0), inference code and router (MIT), and a Gradio demo.
加载中…
点文件 → 加为 tab；按 Esc 关闭