Alexa+bhasha

• 

  alexa-plus-bhasha

Inspiration

Alexa+ launched in India with English, Hindi and Hinglish, and Amazon has said more Indian languages are coming, but without a timeline. That left us thinking about everyone whose home language isn't on the list yet: grandparents, people with low literacy, and families who speak a regional or rare language every day. For them, voice is the easiest way to use technology, and it still doesn't speak their language.

A simple way to see the gap:

$$ \text{Language coverage} = \frac{\text{languages Alexa+ supports}}{\text{languages people actually speak at home}} $$

For rare languages, this ratio is close to zero. We wanted to push it up without waiting for a full rebuild every time.

Alexa+ Bhasha — Every language deserves a voice.

What it does

Alexa+ Bhasha is an MCP-style "change maker" that lets an AI model propose language changes for Alexa+ in a safe, reviewable way:

• A model drafts a change, such as a routine, a response style, or a set of phrases in a new language.
• A human previews exactly what the change would do.
• The change goes live only after explicit approval.
• Any change can be rolled back.

A native speaker of a rare language can help teach Alexa+ their language through reviewed updates, instead of waiting for a large release.

How we built it

• Python and the Model Context Protocol (MCP) SDK (FastMCP) for the server.
• Six tools exposed to the model: list_change_kinds, create_change, list_changes, preview_change, apply_change and rollback_change.
• A draft → confirm → apply → rollback flow. A change is applied only when

$$ \text{apply}(c) = \begin{cases} \text{run} & \text{if } \texttt{confirm} = \text{true} \ \text{reject} & \text{otherwise} \end{cases} $$

• Changes are stored as JSON with status and history, so every action is traceable.
• An AlexaAdapter layer isolates the Alexa+ side. In this version it is simulated, and it is designed to be swapped for real Alexa developer API calls.

[Add here: the rare language you chose for the demo and the commands you showed.]

Challenges we ran into

• No direct Alexa+ access. We couldn't change Alexa+ itself, so we built the adapter as a clean boundary and simulated the apply and rollback steps.
• Safety. Letting a model change a voice assistant is risky, so we made human approval and rollback core to the design instead of an add-on.
• Rare-language data is scarce. There are few datasets and fewer verified speakers, so the design relies on native-speaker review rather than trusting the model alone.
• Scope. Language support is huge, so we focused on a small set of everyday commands for one language to keep the demo honest.

Accomplishments that we're proud of

• A working MCP server with a complete draft, preview, approve and rollback cycle.
• A safety-first design where nothing changes without human confirmation.
• A clear, evidence-based problem: Alexa+ in India covers English, Hindi and Hinglish, and other Indian languages have no announced timeline.
• A design that can grow from one language to many without redesigning the system.

[Add here: anything you actually achieved, such as a working demo in your chosen language or feedback from a native speaker.]

What we learned

• How MCP lets a model call well-defined tools instead of taking free-form actions, which makes AI-driven changes easier to control and audit.
• That the hard part of rare-language support isn't only the model, it's trust: who verifies a language pack, and how to undo a mistake.
• How Alexa+ currently handles languages, and where the real gaps are.
• To design for review first: draft, preview, approve, rollback.

What's next for Alexa+ Bhasha

• Connect the adapter to real Alexa and Alexa+ developer APIs.
• Add a create_language_pack tool and a native-speaker review screen.
• Grow from one language to many, with community contributors verifying each pack.
• Add voice-first onboarding so low-literacy users can contribute and review by speaking.

Built with: Python, MCP (Model Context Protocol), JSON storage

Built With

• accessibility
• alexa
• alexa-plus
• amazon-alexa
• indian-languages
• low-resource-languages
• multilingual
• natural-language-processing
• rare-languages
• voice-ai