Bangla Voice Stack
Self-hosted Bangla speech services
The question
Can production Bangla speech run entirely offline on self-hosted GPUs, down to a Jetson Orin, with a voice cloned from one short reference clip?
Standalone Bangla text-to-speech and speech-to-text microservices on NVIDIA GPUs, with reference-clip voice cloning, streaming transcription, an API gateway with per-client keys, and a SIP call-centre demo. Provides the voice layer for Zayyan OS.
- When
- Jun 2026 – Present
- Where
- Evolution Technologies
- Role
- Creator
- Whose idea
- Mine
- Status
- Ongoing
Sits between
This work crosses speech, edge computing, GPU inference, offline-first, telephony, distributed systems.
- 11 Indic
- Languages
- yes
- Offline
What it is
Two model services and the plumbing around them. The TTS service turns native-script text into speech with a voice defined by a short reference clip and its transcript. The STT service transcribes Bangla with faster-whisper on the GPU or, optionally, Google's streaming API, over HTTP, server-sent events, or a bidirectional WebSocket.
How it works
- Fully offline once the models are downloaded; optional API-key auth and rate limiting on each service.
- An API gateway with per-client keys fronts the speech services and outbound calls.
- A call-centre demo runs GPT-driven personas grounded in uploaded PDF knowledge bases and transliterates stray Latin words into Bengali before synthesis, because the TTS voice has no Latin alphabet.
- Builds for x86 servers and for Jetson Orin (JetPack 6).
Status
In daily use as the speech layer for Zayyan OS.
Around the same time
See 2026 on the timeline →Shares its fields with