vocra_core 0.2.2

Pure-Dart engine behind Vocra: streaming voice-AI orchestration (STT to LLM to TTS) with pluggable Groq, OpenAI, Gemini, xAI, Z.ai, Deepgram, and ElevenLabs providers.

Changelog

0.2.2

Packaging and metadata only — no Dart API or behavior changes.

Added

• A funding link (GitHub Sponsors) and https://www.vocra.cloud in the package metadata.

Fixed

• repository and issue_tracker pointed at the demo app repository, which broke pub.dev's repository verification. Both now resolve to this package's directory in the SDK repository.

Changed

• Licensed under MIT (the 0.2.1 release was published as MIT; an interim proprietary relicense never shipped to pub.dev).

0.2.1

First pub.dev release (developed in-repo as voice_core through 0.1.0).

Added

• Provider facades VocraLlm / VocraTts / VocraStt (e.g. VocraLlm.openAi(apiKey: ...)); new OpenAiLlm + shared OpenAiCompatibleLlm base, plus XaiLlm (Grok) and ZaiLlm (GLM) adapters.
• Typed model/voice catalogs picked by the facades: GroqModel, OpenAiModel, GeminiModel, XaiModel, ZaiModel, DeepgramVoice, ElevenLabsVoice, ElevenLabsModel, DeepgramSttModel — each with values, .custom(id), and a ModelTier; plus LlmVendor/TtsVendor/SttVendor enums for pickers.
• VocraPrompt structured system prompts (sections, json, jsonText) as an alternative to VocraConfig.systemPrompt.
• Conversation control & retrieval: conversation, messages (live aggregated transcript via TranscriptAggregator), mute/unmute/isMuted, endSession(), sessionEnded, lastReport, and the SessionReport model.
• VocraConfig.policies (SessionPolicies: max duration, silence timeout, end phrases, farewell endMessage) and VocraConfig.assistantName.
• Greeting.none(); DeepgramStt language parameter.
• VocraConfig.greeting (Greeting.text / Greeting.generated) so the assistant can speak first when a conversation starts.
• VocraConfig.naturalSpeech: augments the system prompt with a live-voice style guide (brief spoken replies, contractions, no markdown/emojis), plus light audio-tag guidance when the TTS supports tags.
• SpeechTextNormalizer: strips markdown, emojis, and unsupported audio tags from text before it reaches TTS. Transcripts and history keep the original.
• VoiceEngine.speak(text): speak a scripted assistant utterance (no LLM call).
• TtsProvider.supportsAudioTags and LlmProvider.warmUp/TtsProvider.warmUp.
• GeminiLlm (streaming) and ElevenLabsTts providers; ElevenLabsTts gains optional style / useSpeakerBoost voice settings.
• DeepgramStt exposes endpointing / utteranceEnd tuning parameters.

Changed

• Breaking (rename): VoiceConfig → VocraConfig; systemPrompt is now optional (provide it or prompt).
• Breaking (behavior): conversation history now resets on each startConversation() (a session is one conversation).
• Breaking (interface): LlmProvider and TtsProvider gained members (warmUp, supportsAudioTags). Custom providers using implements must add them; the built-in adapters are updated.
• Breaking (default): GroqLlm now defaults to openai/gpt-oss-20b (llama-3.1-8b-instant is retired by Groq on 2026-08-16) and sends reasoning_effort: low / include_reasoning: false for openai/gpt-oss*.
• Text sent to TTS is now normalized (see SpeechTextNormalizer).
• First-turn latency: LLM/TTS network paths are pre-warmed and the mic + STT transport start concurrently at conversation start.

Fixed

• A MicSource.resume() failure after a turn now surfaces as a ProviderError(provider: 'microphone') on the errors stream instead of escaping as an unhandled async exception.

0.1.0

Internal pre-release (git tag v0.1.0, never published): streaming conversation engine (VoiceEngine, TurnMachine, AudioQueue), Groq LLM + Deepgram STT/TTS adapters, and the typed VoiceError hierarchy.