addisai 0.6.0 — Scribe speaker labels

Addis Scribe speaker labels.

• transcribe with backend: "turbo", speakers: true (Python: backend="turbo", speakers=True) labels every word and caption with speaker (1, 2, … in order of first speaking) and returns the total as speakers. Word timestamps are included.
• toSrt() / toVtt() (Python: addisai.to_srt / addisai.to_vtt) add Speaker N: (SRT) or a <v Speaker N> voice span (VTT) to labelled cues.
• Turbo only, on completed uploads. Labels are speaker numbers, not names. Works best with 2 to 4 speakers.

Docs: https://docs.addisassistant.com/docs/capabilities/speech-to-text

addisai 0.5.0 — Scribe timestamps and captions

Addis Scribe word timestamps and captions.

• transcribe with timestamps: "word" (Python: timestamps="word") returns words and segments, which are caption cues timed for display.
• toSrt() / toVtt() (Python: addisai.to_srt / addisai.to_vtt) turn a result into a caption file locally. The output is identical to the API's format=srt|vtt.
• Backends are now standard (default) and turbo.
• Word timestamps work on completed uploads (transcribe), not on stream() or live sessions.

Docs: https://docs.addisassistant.com/docs/capabilities/speech-to-text#timestamps-and-captions

addisai 0.4.0 — Addis Scribe transcription

Adds Addis Scribe Amharic file and live transcription through addis.scribe: uploads, transcript streams, PCM WebSocket connections, session tickets, wallet/rate checks, capability discovery, and request recovery.

Final transcripts include settled wallet usage. Paid uploads and ticket issuance do not automatically retry; account-scoped request IDs support recovery without another charge. The SDKs preserve the existing speech and voice interfaces. Browser apps receive scoped one-use tickets from their server.

Validated with both SDK suites, supported-runtime CI, package builds, and the gateway over HTTP/WebSockets with mocked inference and billing.

addisai 0.3.1 — three-language voice streaming

Voice streaming now consistently covers Amharic, Afaan Oromo, and Tigrinya, with real voice IDs and examples placed within the voice documentation. Requests for other voice languages are rejected before networking.

Realtime sessions can optionally set a maximum generated-audio duration (maxAudioSeconds in JavaScript / max_audio_seconds in Python) when using the upgraded gateway. The field is omitted by default for compatibility with the existing public service. API keys and account JWTs use the SDK's existing authentication headers.

HTTP streaming and persistent WebSocket methods use the same voice pricing and expose completion usage metadata. Keep the same request ID and exact inputs for recovery.

Validation: JavaScript 47 tests and Node 18/20/22 CI passed; Python 44 tests and Python 3.8–3.12 CI passed. Gateway rollout is separate from this SDK release.

addisai 0.3.0 — real-time voice streaming

Addis AI SDK 0.3.0 enables streaming speech through the existing public Voice API.

npm publication pending: the source and built SDK are available in this GitHub release, but npm still serves 0.2.0. The publishing workflow passes validation and is awaiting valid npm publish authorization.

• voice.stream() and textToSpeech.stream() yield MP3 phrases and expose final clip and usage metadata.
• realtime.createSession() and realtime.connect() support persistent voice WebSocket sessions. Browsers can use connectRealtime() with a scoped ticket issued by their application server.
• Helpers support text buffering, cancellation, repeated spoken turns, and stable request IDs for recovery without a second charge. Started synthesis still completes billing after cancellation or disconnect.
• Voice language types now match the catalog; chat, transcription, and translation types include Tigrigna. Their request logic and billing rates are unchanged.
• Node 18 WebSocket support adds ws 8.22.0. The real-time voice guide covers Amharic, Afaan Oromo, and Tigrigna.

Install:

npm install https://github.com/Addis-AI-Org/addisai-js/releases/download/v0.3.0/addisai-0.3.0.tgz

After npm publication completes, use npm install addisai@0.3.0.

Validation: 32 tests, typecheck, and ESM/CJS/declaration builds; CI on Node 18, 20, and 22. Shipped dependencies have zero reported vulnerabilities. Six existing development-tool audit findings remain in the repository. A live test against the shared public voice protocol verified a 0.084 ETB deduction for 1.008 seconds of audio, one ledger charge, and no additional charge for same-ID clip recovery.

This release updates the SDK and does not deploy changes to the gateway or wallet architecture.

addisai 0.2.0

0.2.0 — replaces the retired am-hiwot voice id with the production am-hamen in all examples/docs, preserves Voice 2 minute-pricing fields, and points the homepage at addisassistant.com.

addisai 0.1.2

Docs: corrected the homepage/brand link to https://addisassistant.com (was a placeholder domain). No code changes.

addisai 0.1.1

Build hygiene: source maps are no longer shipped in the npm package (they embedded the full TypeScript source and bloated the tarball ~265kB). A publish-time guard now rejects any build that embeds source in a map. No API or behavior changes.

addisai 0.1.0 — developer preview

First release of the official Addis AI Node SDK: voice (TTS), OpenAI-compatible chat (system/persona/tools), speech-to-text, and translation for Amharic and Afan Oromo. Dual ESM+CJS, typed, zero runtime deps. Persona/function-calling activate as the backend rollout completes.