Multilingualism is not a feature in DAKSH — it is a foundational capability. In a country as linguistically diverse as India, an AI assistant that only supports English or Hindi is fundamentally limited in its reach. DAKSH was built to operate fluently across a spectrum of Indian languages and dialects, with deep understanding and accurate translation not only at the sentence level, but also within specialized enterprise, governance, and customer support contexts. This allows DAKSH to transcend language barriers and serve users in their native linguistic environments with domain-specific accuracy.
a) Language Identification and Switching
Every voice or text query sent to DAKSH first passes through a Language Identification (LID) module. This model determines the primary language of the input based on phonetic cues (for speech) or character tokens (for text). For instance, it can distinguish between Hindi, Marathi, and Chhattisgarhi even when similar vocabulary is used. In code-mixed inputs — where users blend English and a regional language — the system activates dual-tokenizer mode and applies hybrid inference, ensuring that meaning is preserved without misinterpretation.
DAKSH’s multilingual engine is not language-agnostic; it is language-aware, meaning it adapts its tokenization, embeddings, and model prompts based on the identified language. This provides significantly higher fidelity than generic multilingual models which often treat languages as equivalent across vocabulary and syntax.
b) Context Preservation Across Languages
A key challenge in multilingual systems is context preservation, especially in multi-turn conversations. Users may begin a conversation in English, switch to Hindi midway, and finish in their regional dialect. DAKSH handles this by:
-
Maintaining conversation state across different language segments using internally normalized vector representations.
-
Tracking slot-filling intents (like name, date, place) regardless of the language used to express them.
-
Translating knowledgebase chunks retrieved during RAG so that the response language always matches the query language.
This allows a Marathi-speaking user to ask a policy question and receive an answer pulled from a Hindi document — seamlessly translated and rephrased in Marathi with full fidelity.
c) Language-Specific Model Routing
Instead of using a single universal model for all languages, DAKSH employs language-specific pipelines when required. This includes:
-
Custom tokenizers with regional vocabulary
-
Fine-tuned embedding spaces that retain linguistic nuances
-
Language-conditioned prompts that adapt grammar and phrasing
This design ensures that translated content doesn’t just sound linguistically accurate but also culturally contextual. For example, a deadline notification in Tamil is not just converted word-for-word from English — it is localized in tone, date format, and courtesy level based on regional expectations.
d) Domain-Adaptive Multilingual Training
DAKSH’s multilingual capability stems not just from general-purpose corpora but from domain-specific bilingual data. During model training and fine-tuning, the following techniques were applied:
-
Sentence-level and paragraph-level alignment of bilingual enterprise documents
-
QA pairs annotated in both source and target languages with identical semantic intent
-
Training on code-mixed chat logs to enhance conversational adaptability
The result is that DAKSH does not “translate” in the traditional sense — it “understands” in every supported language. This native comprehension enables it to resolve disambiguations (e.g., "form" as document vs. physical shape) and recognize honorifics, idioms, and implied meanings across diverse linguistic structures.
e) Dynamic Output Language Control
DAKSH offers both automatic and user-controlled output language selection. Users can:
-
Speak or type a query in any supported language and receive a response in the same
-
Choose a preferred output language from a profile setting or dropdown
-
Switch language mid-conversation without loss of session context
In institutional deployments, administrators can configure default languages by region — ensuring that voice kiosks in Tamil Nadu default to Tamil, while dashboards in Maharashtra speak Marathi by default.