prathoshap/vagdhenu

Vāgdhenu — metered Sanskrit/Vedic chant text-to-speech (DiT + BigVGAN). Apache-2.0.

What it solves

Vágdhenu is a production-grade text-to-speech (TTS) system specifically designed for Sanskrit chanting (páráyaána). Unlike standard TTS, which often reads text flatly, this system generates audio with metrically-aware durations and melodic contours faithful to traditional chanting styles.

How it works

The system uses a flow-matching Diffusion Transformer (DiT) backbone based on IndicF5/F5-TTS. To avoid issues like Hindi schwa-deletion, Sanskrit text is routed through the Kannada script. It employs a fine-tuned NVIDIA BigVGAN-v2 vocoder to ensure stability on long vowels.

Prosody is managed through a combination of a reference clip (controlling voice, swara, and pace) and voice-steering fine-tuning. The system also includes a specialized text frontend that handles complex Sanskrit linguistic rules, including visarga sandhi, homorganic anusvára, and meter/gaána detection.

Who it’s for

It is designed for researchers, developers, and practitioners interested in the synthesis of sacred Sanskrit recitation for study, accessibility, and the production of chanted audio content.

Highlights

  • High Fidelity: Achieves a Mean Opinion Score (MOS) of ~4.6 from expert listeners.
  • Linguistic Precision: Correctly renders complex conjuncts, including retroflex aspirates.
  • Production Proven: Used to generate over 17 hours of audio for the Mahábhárata Tátparya Niráya and over 16,000 verses for the ŜśƱmad Bhágavatam.
  • Advanced Frontend: Includes a custom text-processing pipeline for Devanagari to Kannada routing and meter detection.

Related

  • Project
    SWivid/F5-TTS

    F5-TTS is a high-performance text-to-speech system using a Diffusion Transformer and flow matching to generate fluent, faithful speech with fast training and inference.

  • Project
    MoonInTheRiver/DiffSinger

    A diffusion-based framework for high-quality singing voice synthesis and text-to-speech generation from lyrics, MIDI, and pitch data.

  • Project
    pnnbao97/VieNeu-TTS

    VieNeu‑TTS is an open‑source, on‑device Vietnamese text‑to‑speech system (v3 Turbo) that runs on CPU via ONNX (or CUDA via PyTorch for large batches). It offers 48 kHz high‑fidelity speech, 23 preset voices, instant voice cloning from a 3‑8 s clip, emotion tags, batch/conversation mode, and a low‑latency streaming API. An edge‑device variant (v3 Nano) provides faster inference on weak CPUs at reduced quality. The project includes a Gradio web UI and a Python SDK (`vieneu`). A newer v4 model is proprietary and only available through the hosted API.

  • Project
    VOICEVOX/voicevox_engine

    VOICEVOX Engine is an open‑source HTTP server that wraps the VOICEVOX Japanese TTS core. It lets you send text (or a detailed synthesis query) via REST endpoints and receive WAV audio, with support for streaming, user dictionaries, presets, voice morphing, singing synthesis, and optional GPU acceleration. Docker images and a mock mode make deployment easy, and the API is documented at /docs.

  • Project
    tronghieuit/v-tts

    A lightweight Vietnamese text-to-speech system featuring multi-speaker synthesis and zero-shot voice cloning that runs efficiently on CPU.