What’s new. Shipped from Dar es Salaam.

Our official SDKs are out in beta.

• Python: pip install waridi. Sync and async clients, Python 3.9 and newer.
• Node.js / TypeScript: npm install @waridi/sdk. Fully typed, Node 18 and newer, plus @waridi/sdk/web for browser calls.
• Both cover text-to-speech, speech-to-text, the streaming WebSocket, agents, voice cloning and dubbing.
• Automatic retries with backoff on 429 and 5xx responses, and typed errors that carry the request_id.

Every code sample in the docs now has Python, TypeScript and cURL tabs. Start with the quickstart.

These are beta releases, so method names may still change before 1.0. We will flag any change here first.

You can now build Waridi Agents in the sandbox: voice agents that answer in Swahili, understand callers who switch to English, and hand over to a person when needed.

• POST /v1/agents creates an agent with a voice, greeting, instructions and tool webhooks.
• Each agent gets a SIP URI. Point a test trunk at it and call in. See Telephony.
• Browser and app calls over WebRTC, using short-lived client secrets.
• Callers can interrupt the agent mid-sentence, and it stops and listens.

Our latency target for agents is first reply audio in about 500 ms, and always under 800 ms. This is a target we are working toward during the sandbox, not a measured result.

Agents stay in the sandbox for now. If you would like to pilot one on a real line, write to hello@waridi.io. Start with the voice agents guide.

Waridi Sikia, our speech recognition model, is now in beta.

• POST /v1/stt transcribes files with sikia-1, including segment and word timestamps.
• Code-switching is on by default. Each segment is tagged sw, en or sw-en.
• New streaming API at wss://api.waridi.io/v1/stream with partial and final transcripts, voice activity events and 8 kHz μ-law input for phone audio.
• Speaker diarization is available as an opt-in beta.

What we are still working on: Sheng accuracy is experimental, and very noisy market or matatu recordings need more work. We are building a public Swahili ASR benchmark and will publish results with methodology once it is ready.

The first preview of Waridi Sauti is live for everyone in early access.

• POST /v1/tts with two models: sauti-standard for fast, clear prompts and sauti-pro for more expressive delivery.
• Six preview voices: zawadi, baraka, neema and juma (Tanzanian), wanjiru and otieno (Kenyan).
• language accepts sw-TZ (default) and sw-KE, and changes how numbers, currency and times are read.
• Output as mp3, wav, opus, or mulaw at 8 kHz for phone lines.
• Mixed Swahili–English sentences are pronounced naturally, with no special markup needed.

Known gaps: long paragraphs on sauti-standard can sound flat, and some place names outside Dar es Salaam are mispronounced. Use the pronunciations field for now and send us the words you hit at hello@waridi.io.

Read the text-to-speech guide.

Today we opened Waridi to our first group of early-access developers. It is small on purpose. We want to hear from people building real Swahili voice products before we open the doors wider.

What is available:

• API keys and the Free sandbox, with 10 minutes of audio per month.
• Waridi Studio, the web app for trying voices and transcribing files without code.
• The base URL https://api.waridi.io/v1 and bearer-key authentication.
• First drafts of these docs.

What is not ready yet: speech generation and recognition arrive as separate previews over the coming weeks. We will post each one here.

Everything in early access can change. When something breaks compatibility, we will say so in this changelog at least 14 days ahead. Request access at waridi.io/get-started.