Vagbodha - brahm-ai-official/brahm-ai GitHub Wiki

Vāgbodha

Vāgbodha is the multilingual voice intelligence, neural speech, translation, pronunciation, mantra-recitation, and audio-production system of the BRAHM-AI ecosystem.

It was developed to provide a common voice layer for BRAHM AI Acharya, BRAHM LOK, and other BRAHM-AI experiences while maintaining native language voices, specialized Sanskrit processing, local translation, pronunciation intelligence, and provider-independent public voice identities.

Vāgbodha is not designed as a conventional Text-to-Speech textbox. Its architecture separates language detection, translation, pronunciation preparation, Sanskrit processing, mantra processing, voice selection, speech synthesis, and audio management into distinct layers.

Official service:
https://brahm-ai.in/vagbodha/

Overview

Field Description
Name Vāgbodha
Developer BRAHM-AI
Category Multilingual Voice Intelligence
Primary functions Neural speech, translation, pronunciation, Sanskrit and mantra recitation, audio projects
Core systems BRAHM Local Voice Engine, BRAHM Local Translation Engine
Language architecture Native language-specific neural voice profiles
Specialized Sanskrit system Sanskrit, Shloka, Mantra and Stotra processing
Mantra voices Ritvik (Male), Shruti (Female)
General voice architecture Male and female native neural profiles per supported language
Connected systems BRAHM AI Acharya, BRAHM LOK, BRAHM Voice
Processing direction Local-first, modular and provider-independent
Interface Web-based voice and audio workspace
Status Public Beta / Active development
Official website https://brahm-ai.in/vagbodha/

Etymology

The name Vāgbodha combines two Sanskrit concepts:

  • Vāk / Vāg (वाक्) — speech, voice or expression
  • Bodha (बोध) — understanding, awareness or knowledge

The name represents the system's objective of combining language understanding with spoken expression.

Purpose

Vāgbodha was developed to provide a unified BRAHM-AI voice layer instead of building an unrelated speech implementation for every product.

Its functions include:

  • multilingual neural speech;
  • native male and female voices;
  • Indian-language speech;
  • international-language speech;
  • Sanskrit reading;
  • mantra recitation;
  • shloka reading;
  • stotra reading;
  • pronunciation preparation;
  • local translation;
  • mixed-language processing;
  • study audio;
  • creative narration;
  • character speech;
  • game narration;
  • knowledge-response audio;
  • project-based audio generation.

The architecture is intended to allow the same voice intelligence to be reused across multiple BRAHM-AI systems.

Development

Vāgbodha reached a working runtime stage during August 2026.

Development initially focused on:

  • Indian-language neural speech;
  • Sanskrit-aware text handling;
  • pronunciation preparation;
  • male and female voice selection;
  • study audio;
  • creative narration;
  • shloka and mantra speech;
  • project-based audio management.

The architecture was subsequently expanded with:

  • BRAHM Local Voice Engine;
  • BRAHM Local Translation Engine;
  • international languages;
  • additional native neural voice profiles;
  • BRAHM AI Acharya integration;
  • BRAHM LOK integration;
  • specialized mantra speech processing.

Core Architecture

Vāgbodha uses a modular processing pipeline.

                         USER TEXT
                             │
                             ▼
                 Language & Script Detection
                             │
                             ▼
                    Content Classification
                             │
            ┌────────────────┼────────────────┐
            │                │                │
            ▼                ▼                ▼
         General          Indic /         Mantra /
          Speech          Sanskrit         Shloka
            │                │                │
            └────────────────┼────────────────┘
                             │
                             ▼
                  Translation if Required
                             │
                             ▼
                 Pronunciation Preparation
                             │
                             ▼
                    Native Voice Router
                             │
               ┌─────────────┴─────────────┐
               ▼                           ▼
          Male Profile                Female Profile
               │                           │
               └─────────────┬─────────────┘
                             ▼
                  BRAHM Local Voice Engine
                             │
                             ▼
                 Neural / Supported Runtime
                             │
                             ▼
                       Audio Output
                             │
                             ▼
                   Project Audio Library

Translation, pronunciation and voice generation remain logically independent.

BRAHM Local Voice Engine

The BRAHM Local Voice Engine is the voice-generation and routing architecture developed for Vāgbodha and connected BRAHM-AI products.

Its responsibilities include:

  • text preparation;
  • language identification;
  • native voice selection;
  • male/female voice routing;
  • pronunciation handling;
  • Sanskrit segmentation;
  • mantra-mode routing;
  • speech pacing;
  • long-form segmentation;
  • neural speech generation;
  • supported local inference;
  • fallback management;
  • audio processing.

The Local Voice Engine allows BRAHM-AI to maintain its own voice architecture while changing or extending underlying speech infrastructure when required.

Provider-Independent Voice Identity

Vāgbodha does not require users to select infrastructure-provider voice identifiers.

Instead, the public system uses BRAHM-managed voice identities.

Public BRAHM Voice Identity
             ↓
      Vāgbodha Voice Router
             ↓
       Language Profile
             ↓
    BRAHM Local Voice Engine
             ↓
 Local / Neural / Supported Path
             ↓
            Audio

This separation allows:

  • infrastructure replacement;
  • provider fallback;
  • local inference;
  • voice-model upgrades;
  • language-specific optimization

without changing the public voice identity.

Native Neural Voice Architecture

Vāgbodha uses language-specific neural voice profiles.

A Bengali voice is therefore treated as a Bengali voice path, a Nepali voice as a Nepali voice path, an Odia voice as an Odia voice path, and an Assamese voice as an Assamese voice path.

Languages may share parts of the wider runtime architecture, but they remain separate language identities at the Vāgbodha layer.

Most supported languages provide both:

  • Male neural voice
  • Female neural voice

The public names below are BRAHM voice identities and do not expose internal provider voice identifiers.

Indian & Indic Voice Identities

Language Native Name Male Voice Female Voice
Hindi हिन्दी Ojas Ira
Sanskrit संस्कृतम् Ritvik Shruti
Bengali বাংলা Arin Nandini
Assamese অসমীয়া Neel Jonaki
Odia ଓଡ଼ିଆ Anshuman Subhashree
Nepali नेपाली Himal Prakriti
Marathi मराठी Soham Revati
Gujarati ગુજરાતી Mihir Dhara
Punjabi ਪੰਜਾਬੀ Arjan Heer
Tamil தமிழ் Kavin Yazhini
Telugu తెలుగు Vihan Sravani
Kannada ಕನ್ನಡ Kiran Anvitha
Malayalam മലയാളം Nilan Meera
Indian English English Ayaan Maya

These names represent the public BRAHM identity layer.

They are not the names of underlying providers or infrastructure models.

International Voice Identities

Vāgbodha also provides native-language neural voice paths for supported international languages.

Language Native Name Male Voice Female Voice
French Français Luc Claire
Russian Русский Mikhail Anya
Uzbek O‘zbekcha Timur Dilnoza
Polish Polski Jan Zofia
Italian Italiano Luca Giulia
Indonesian Bahasa Indonesia Arif Sari
Spanish Español Mateo Lucía
Chinese 中文 Wei Lian
Japanese 日本語 Ren Aoi

Each language is routed through an appropriate native-language speech profile rather than forcing all languages through one universal voice.

Common Voice Identity Layer

The Vāgbodha public identity layer can therefore be represented as:

                  VĀGBODHA VOICES

                       Text
                        │
                        ▼
                 Language Router
                        │
       ┌────────────────┼─────────────────┐
       │                │                 │
       ▼                ▼                 ▼
     Hindi           Bengali          Japanese
  Ojas / Ira       Arin / Nandini      Ren / Aoi

       │                │                 │
       └────────────────┼─────────────────┘
                        ▼
               Native Neural Runtime
                        │
                        ▼
                       Audio

The same principle applies to the other supported languages.

Sanskrit Voice System

Sanskrit requires additional processing compared with normal conversational speech.

Vāgbodha therefore provides a dedicated Sanskrit-aware processing path.

Supported material includes:

  • Sanskrit prose;
  • shlokas;
  • mantras;
  • stotras;
  • prayers;
  • scripture verses;
  • dhyana passages;
  • philosophical passages;
  • mixed Sanskrit and explanation;
  • repeated recitation.

Sanskrit Public Voices

The Sanskrit voice layer includes:

Ritvik

Ritvik is the male Sanskrit and mantra-oriented BRAHM voice identity.

Its intended use includes:

  • Sanskrit passages;
  • mantra;
  • shloka;
  • stotra;
  • invocation;
  • sacred-text recitation.

Shruti

Shruti is the female Sanskrit and mantra-oriented BRAHM voice identity.

Its intended use includes:

  • Sanskrit passages;
  • mantra;
  • shloka;
  • stotra;
  • prayers;
  • scripture reading;
  • meditative recitation.

Ritvik and Shruti are separate from ordinary conversational voice identities.

BRAHM Mantra Voice System

Vāgbodha contains a specialized Mantra Voice System.

The system exists because mantra and Sanskrit verse should not always be processed with the same prosody used for normal prose.

The Mantra Voice System handles content such as:

  • mantra;
  • shloka;
  • stotra;
  • Sanskrit prayer;
  • invocation;
  • scripture recitation;
  • japa-oriented repeated text;
  • dhyana verses;
  • traditional verse structures.

Mantra Voice Pipeline

Canonical Sanskrit
        │
        ▼
Sanskrit Detection
        │
        ▼
Mantra / Shloka Classification
        │
        ▼
Text Preservation
        │
        ▼
Pronunciation Preparation
        │
        ▼
Verse & Phrase Segmentation
        │
        ▼
Cadence / Pause Planning
        │
        ▼
 ┌───────────────┬───────────────┐
 ▼                               ▼
Ritvik                         Shruti
Male Mantra Voice          Female Mantra Voice
 │                               │
 └───────────────┬───────────────┘
                 ▼
          Neural Recitation
                 │
                 ▼
               Audio

Mantra-Specific Processing

The mantra-processing layer can account for:

  • danda ।;
  • double danda ॥;
  • verse boundaries;
  • phrase boundaries;
  • Sanskrit phonetic structure;
  • consonant clusters;
  • long and short vowel behavior;
  • anusvara;
  • visarga;
  • pronunciation-sensitive compounds;
  • controlled pauses;
  • recitation cadence;
  • repeated mantra phrases.

The objective is not simply to slow ordinary TTS.

The text first passes through a Sanskrit and pronunciation preparation stage before speech generation.

Vedic Accent Limitation

Vāgbodha distinguishes ordinary Sanskrit/mantra pronunciation from verified Vedic svara.

Where canonical accent information is not available, Vāgbodha does not present generated pitch behaviour as authenticated Vedic recitation.

This distinction is maintained between:

  • Sanskrit pronunciation;
  • mantra-style recitation;
  • Vedic svara;
  • historically verified chant traditions.

Pronunciation Guard

Vāgbodha includes a Pronunciation Guard layer for pronunciation-sensitive text.

It can process:

  • Sanskrit;
  • Indian-language text;
  • mixed scripts;
  • proper names;
  • scripture terminology;
  • numbers;
  • punctuation;
  • abbreviations;
  • sentence boundaries;
  • voice-friendly normalization.
Original Text
      │
      ▼
Protected Source
      │
      ▼
Pronunciation Analysis
      │
      ▼
Speech Representation
      │
      ▼
Voice Engine

The original text and the speech-preparation representation remain separate.

Canonical Text Protection

When Vāgbodha is used with scripture, the speech engine should not silently modify canonical source text.

Canonical Source
      │
      ├──────────────► Displayed / Preserved
      │
      ▼
Speech Preparation Copy
      │
      ▼
Pronunciation Engine
      │
      ▼
Audio

This distinction is particularly important for:

  • Sanskrit;
  • Vedas;
  • Bhagavad Gita;
  • Ramayan;
  • mantras;
  • Thirukkural;
  • other canonical texts.

BRAHM Local Translation Engine

The BRAHM Local Translation Engine provides the translation layer used by Vāgbodha and other supported BRAHM-AI workflows.

It is logically independent from speech generation.

Vāgbodha can therefore operate in three primary configurations.

Translation Only

Source Text
     ↓
Language Detection
     ↓
BRAHM Local Translation Engine
     ↓
Target-Language Text

Voice Only

Source Text
     ↓
Pronunciation Preparation
     ↓
Native Voice Selection
     ↓
BRAHM Local Voice Engine
     ↓
Audio

Translation + Voice

Source Text
     ↓
Source-Language Detection
     ↓
BRAHM Local Translation Engine
     ↓
Target-Language Text
     ↓
Target-Language Pronunciation
     ↓
Native Voice
     ↓
Audio

Translation Architecture

The translation layer is designed for:

  • Indian languages;
  • international languages;
  • mixed-language input;
  • script-sensitive translation;
  • local/private processing;
  • voice workflows;
  • source-text preservation;
  • reduced dependency on a single external translation service.

Translation does not alter protected canonical source material unless translation itself is explicitly requested.

Language Detection

The Vāgbodha language layer identifies or receives the intended language before voice generation.

The processing profile can use:

  • language selection;
  • script;
  • detected language;
  • mixed-language segments;
  • content classification;
  • Sanskrit detection;
  • target language.
Input
  ↓
Language
  ↓
Script
  ↓
Content Type
  ↓
Processing Profile

Explicit user language selection can take priority where automatic script detection would be ambiguous.

Mixed-Language Speech

Vāgbodha supports mixed-language content.

Examples include:

  • Hindi + English;
  • Bengali + English;
  • Tamil + English;
  • Sanskrit + Hindi;
  • Sanskrit + English;
  • multilingual study material.

The speech planner may divide the text into language-aware segments.

Multilingual Text
       ↓
Segment Detection
       ↓
Language Classification
       ↓
Segment-Specific Preparation
       ↓
Native Voice Planning
       ↓
Unified Audio

Voice Modes

The public Vāgbodha interface organizes generation around three primary modes.

Study

Study Mode is intended for:

  • notes;
  • lessons;
  • revision material;
  • explanations;
  • educational content;
  • scripture study;
  • long-form learning audio.

The mode prioritizes clarity and sustained listening.

Create

Create Mode is intended for:

  • narration;
  • stories;
  • scripts;
  • media;
  • creative audio;
  • character dialogue;
  • spoken content.

Indic

Indic Mode is intended for pronunciation-sensitive Indian-language material.

Typical content includes:

  • Sanskrit;
  • shlokas;
  • mantras;
  • stotras;
  • Indian-language literature;
  • traditional terminology;
  • mixed Sanskrit passages.

Indic Mode can invoke the specialized Sanskrit and mantra-processing layers when required.

Smart Routing

Vāgbodha can determine an appropriate processing path according to the input.

Input Text
    ↓
Smart Analysis
    ↓
 ┌───────────┬─────────────┬────────────┐
 ▼           ▼             ▼            ▼
Study      Create         Indic       Mantra
 │           │             │            │
 └───────────┴──────┬──────┴────────────┘
                    ▼
               Voice Router

This reduces the need for users to understand the internal speech architecture.

Speaking Controls

Supported workflows can use settings such as:

  • language;
  • male/female voice;
  • voice identity;
  • mode;
  • speed;
  • pronunciation profile;
  • content type.

These controls remain independent of the underlying provider configuration.

Project Audio

Vāgbodha includes project-based audio organization.

Generated speech can be organized through:

  • Recent Projects;
  • Recent Voices;
  • generated-audio history;
  • project grouping;
  • repeat access;
  • audio playback;
  • future long-form project workflows.

This makes Vāgbodha an audio workspace rather than a disposable TTS request interface.

Relationship with BRAHM AI Acharya

BRAHM AI Acharya uses Vāgbodha as its connected multilingual voice layer.

Acharya is responsible for:

  • Sanskrit intelligence;
  • scripture reasoning;
  • source verification;
  • contextual explanation;
  • grammar;
  • knowledge presentation.

Vāgbodha is responsible for:

  • language-aware speech;
  • native neural voices;
  • pronunciation preparation;
  • Sanskrit speech;
  • mantra speech;
  • multilingual audio.
User
  ↓
BRAHM AI Acharya
  ↓
Knowledge / Verified Response
  ↓
Language Presentation
  ↓
Vāgbodha
  ↓
Native Voice / Mantra Voice
  ↓
Audio

The supported multilingual voice identities are available to the Acharya experience through the shared BRAHM voice architecture.

Acharya therefore does not require an independent TTS system for every language.

See:

BRAHM AI Acharya

Acharya Mantra Integration

Acharya explicitly handles:

  • शास्त्र;
  • संस्कृत;
  • मंत्र;
  • व्याकरण;
  • उच्चारण.

For mantra-oriented content, the architecture can route the Sanskrit block separately from the explanation.

Acharya Response
      │
      ├── Sanskrit / Mantra
      │          ↓
      │   Mantra Voice System
      │    Ritvik / Shruti
      │
      └── Explanation
                 ↓
         User-Language Voice

This allows the mantra to retain a Sanskrit-specific voice treatment while the explanation is spoken in the user's selected language.

Relationship with BRAHM LOK

The same Vāgbodha multilingual voice architecture is integrated with BRAHM LOK.

It can support:

  • BRAHM AI Acharya inside the world;
  • character dialogue;
  • knowledge interactions;
  • multilingual NPC speech;
  • Akashvani;
  • quest dialogue;
  • narration;
  • world events;
  • temple-linked speech;
  • player-language responses.
BRAHM LOK Event / Character
          ↓
    Dialogue or Knowledge
          ↓
      Language Router
          ↓
        Vāgbodha
          ↓
 Native Language Voice Profile
          ↓
      In-World Audio

The objective is for a user to interact with the BRAHM LOK world in the same supported language environment used across Acharya.

See:

BRAHM LOK

Akashvani

Akashvani is a special world-voice role within BRAHM LOK.

It may be used for:

  • world announcements;
  • location narration;
  • knowledge events;
  • quest information;
  • spiritual or symbolic moments;
  • world-system messages.

Akashvani uses the broader Vāgbodha voice architecture while remaining an independent in-world voice role.

Character Voice Architecture

BRAHM LOK can map characters to language-aware public voice identities.

Character
    ↓
Character Voice Role
    ↓
Player Language
    ↓
Vāgbodha Voice Profile
    ↓
Neural Speech

This enables characters to retain their identity while speaking in supported user languages.

Relationship with BRAHM Voice

BRAHM Voice and Vāgbodha serve different purposes.

System Role
BRAHM Voice Real-time conversational voice and avatar interaction
Vāgbodha Voice intelligence, translation, pronunciation and neural audio
BRAHM AI Acharya Knowledge, Sanskrit and scripture intelligence
BRAHM LOK Interactive virtual-world experience

Vāgbodha technologies can be reused across these systems without requiring each product to create an independent voice architecture.

Relationship with Garudaa-LLMG-108

Garudaa-LLMG-108 is BRAHM-AI's own 8B fine-tuned model for Hindi-first, Indian-language, Devanagari, Sanskrit-aware, multilingual and local intelligence workflows.

Garudaa and Vāgbodha perform different roles.

Garudaa / BRAHM Intelligence
            ↓
        Response Text
            ↓
         Vāgbodha
            ↓
     Language & Pronunciation
            ↓
        Neural Voice

Garudaa performs language reasoning where appropriate.

Vāgbodha converts prepared language output into speech and also provides translation workflows.

See:

Garudaa-LLMG-108

Local-First Design

Vāgbodha follows a local-first architectural direction.

Local-first does not require every generation to execute through exactly one local model.

Instead, it means BRAHM-AI maintains control over:

  • voice routing;
  • translation routing;
  • language processing;
  • pronunciation preparation;
  • voice identity;
  • supported local processing;
  • fallback selection.

This allows the runtime to evolve without changing the public product architecture.

Native Voice Principle

A central design principle is:

Use the appropriate language voice rather than making one voice imitate every language.

This is why Vāgbodha maintains separate voice identities for languages such as:

  • Bengali;
  • Assamese;
  • Odia;
  • Nepali;
  • Tamil;
  • Telugu;
  • Kannada;
  • Malayalam;
  • Marathi;
  • Gujarati;
  • Punjabi;
  • French;
  • Russian;
  • Japanese;
  • Chinese;
  • and other supported languages.

Native pronunciation characteristics should be handled by the corresponding language voice path wherever supported.

Privacy

A local or privately controlled voice architecture can reduce unnecessary external data transfer.

Vāgbodha's design supports future deployment models including:

  • local inference;
  • private inference;
  • server inference;
  • hybrid inference;
  • institutional deployment;
  • API-based deployment.

Actual execution may depend on:

  • language;
  • voice;
  • hardware;
  • latency;
  • model availability;
  • deployment environment.

Protected Components

Public Vāgbodha documentation intentionally does not expose:

  • underlying provider voice IDs;
  • private voice models;
  • model weights;
  • voice-model training data;
  • translation-model internals;
  • private fine-tuning data;
  • proprietary pronunciation rules;
  • protected prompts;
  • credentials;
  • internal deployment paths;
  • production secrets.

Public BRAHM voice names represent the product-facing voice identity.

Reliability Principles

Vāgbodha follows several technical principles:

  • preserve original user text;
  • preserve canonical Sanskrit;
  • separate translation from original source;
  • separate speech preparation from stored text;
  • use native language voice paths;
  • distinguish prose from mantra;
  • use pronunciation-aware processing;
  • keep public voice identity independent from infrastructure;
  • allow supported fallback;
  • avoid presenting generated transformations as canonical text.

Limitations

Vāgbodha remains under active development.

Potential limitations include:

  • pronunciation variation;
  • regional accent variation;
  • translation ambiguity;
  • mixed-language ambiguity;
  • infrastructure-dependent latency;
  • long-form generation limits;
  • differences between language voice capabilities;
  • incomplete local inference coverage;
  • varying support across devices.

A language being supported by Acharya does not necessarily imply that every speech feature behaves identically across all languages.

Current Ecosystem Role

Vāgbodha now functions as a shared speech and language layer across the BRAHM-AI ecosystem.

                     BRAHM-AI
                        │
          ┌─────────────┼─────────────┐
          │             │             │
          ▼             ▼             ▼
      Acharya       BRAHM LOK     BRAHM Voice
          │             │             │
          └─────────────┼─────────────┘
                        ▼
                     Vāgbodha
                        │
        ┌───────────────┼────────────────┐
        ▼               ▼                ▼
 Local Translation  Pronunciation    Voice Engine
        │               │                │
        └───────────────┼────────────────┘
                        ▼
              Native Neural Voices
                        │
                        ▼
                      Audio

Future Development

Ongoing Vāgbodha development includes:

  • improved local neural voice quality;
  • stronger native pronunciation;
  • expanded language coverage;
  • improved Sanskrit recitation;
  • improved mantra cadence;
  • long-form audio;
  • streaming speech;
  • lower latency;
  • character voice profiles;
  • BRAHM LOK world voices;
  • additional Acharya voice integration;
  • private deployment;
  • local deployment;
  • Developer API access;
  • expanded project audio;
  • additional BRAHM voice identities.

Related Systems

Official Links

See Also

BRAHM-AI — Ancient Knowledge. Modern Intelligence. Connected Action.

“सत्यं शिवं सुन्दरम्”

© 2025–2026 BRAHM-AI and Royal Web 3 Blockchain. All Rights Reserved.