🎙️ VoicEra

Voice AI is increasingly powerful, but production deployments can create lock-in around models, telephony, data, and infrastructure.

VoicEra is an open infrastructure layer that puts those choices back with the operator.

• Self-hosted by design — run the platform on infrastructure you control.
• Provider neutral — swap STT, TTS, LLM, and telephony providers without rewriting the platform.
• Data ownership — call media, transcripts, and recordings stay in your infrastructure.
• Public-good friendly — Apache 2.0 licensed, transparent, forkable, and deployable without a VoicEra-managed service.
• Composable — use cloud APIs, local models, or a mix of both.

VoicEra is platform you own, not another AI vendor.

Prerequisite: Docker must be installed and running before you start. Launch Docker Desktop (or your Docker daemon) first — the script fails if it can't reach the Docker daemon.

git clone https://github.com/COSS-India/VoicEra.git
cd VoicEra

./scripts/start-application-services.sh $(ARGS)

./scripts/start-application-services.sh $(ARGS) creates the environment, generates required secrets, and starts the stack.

Once running:

./scripts/stop-application-services.sh $(ARGS)

See the documentation for production deployment and configuration.

You bring the models and telephony account. VoicEra connects them into a deployable system.

flowchart LR
  C["Caller"] <--> T["Telephony<br/>Vobiz · Plivo"]

  T -->|"audio"| R["Voice Runtime<br/>Pipecat"]
  R <--> P["AI Providers<br/>STT · LLM · TTS"]

  U["Operator"] --> D["Dashboard"]
  D --> A["API"]
  A --> R

  A --> DB[("FerretDB<br/>PostgreSQL")]
  A --> S[("MinIO<br/>Media & artifacts")]
  A --> Q[("Redis")]
  Q --> W["Workers"]

  R --> S
  W --> DB

The platform separates control plane from real-time execution:

• API — agents, configuration, campaigns, authentication, and orchestration.
• Runtime — live call audio and model interaction.
• Dashboard — operator interface.
• Workers — asynchronous jobs and campaign execution.
• Storage — self-hosted database, object storage, and queue.
• Providers — interchangeable AI and telephony integrations.

Read the architecture guide →

VoicEra follows principles that matter for Digital Public Goods:

VoicEra is intended to be reused, adapted, and independently operated — including by governments, NGOs, public-interest organisations, and other open-source projects.

Integrations live behind stable interfaces, so adding a provider should not require changing the core runtime.

Create a provider under:

apps/providers/{cloud,adapters,local}/

STT, TTS, and LLM providers register through the provider registry.

Add an AI provider →

Implement the telephony contract under:

apps/telephony/providers/

Existing Vobiz and Plivo integrations provide reference implementations.

Add a telephony provider →

Use the optional model server to expose self-hosted STT, TTS, or LLMs through a common gateway.

Agent
  ↓
VoicEra
  ↓
Model Server
  ├── STT
  ├── TTS
  └── LLM

VoicEra is designed to make voice AI more accessible for low-resource and underserved languages — where commercial models, tooling, and high-quality training data are often limited.

It brings together open and interoperable integrations across speech and language technologies, including:

• Bhashini — Indian-language STT and TTS
• AI4Bharat — Indic speech and language models
• Kenpath Vistaar — LLM support for underserved languages
• Cloud providers — additional STT, TTS, and LLM options

The architecture makes it possible to combine these models, self-host them, or replace them as better language technologies emerge.

See the provider registry for the current list.

VoicEra/
├── apps/
│   ├── api/            FastAPI control plane
│   ├── runtime/        Pipecat real-time voice runtime
│   ├── providers/      STT · TTS · LLM integrations
│   └── telephony/      Telephony integrations
├── frontend/           Next.js dashboard
├── model-server/       Optional self-hosted model gateway
├── scripts/             Service lifecycle scripts
└── docs/                Mintlify documentation

VoicEra does not provide telephony accounts or AI inference.

For a real phone deployment, you need:

1. STT, TTS, and LLM access — cloud providers or self-hosted models.
2. A telephony provider — currently Vobiz or Plivo.
3. Infrastructure to run VoicEra and store your data.

For browser-based testing, telephony is not required.

• Getting started
• Architecture
• Provider registry
• Environment variables
• Model server
• Operator FAQ

VoicEra is built in the open.

Use → Adapt → Integrate → Contribute

Before opening a pull request, read CONTRIBUTING.md.

Security issues should be reported privately according to SECURITY.md.

VoicEra is released under the Apache License 2.0.