@animalabs/voice-kit

A provider seam for speech in Connectome. It defines one set of STT and TTS interfaces, adapters for the vendors we use, and the audio helpers around them. It knows nothing about Discord or any other surface: consumers such as @animalabs/discord-mcpl's voice leg open sessions and streams through these interfaces and never touch a vendor API directly.

npm install @animalabs/voice-kit

What's in it

| Export | What it is | |---|---| | SttProvider, SttSession, SttTranscript, SttSessionOptions | Speech-to-text interfaces | | TtsProvider, TtsStream, TtsVoice, TtsAlignment | Text-to-speech interfaces | | ScribeSttProvider | ElevenLabs Scribe realtime STT | | AssemblyAiSttProvider | AssemblyAI realtime STT | | ElevenLabsTtsProvider | ElevenLabs streaming TTS (stream-input), with character timing | | EnergyVad, dbfs | An energy-based voice activity detector | | downmixStereoToMono, monoTo48kStereo | PCM conversion to and from Discord's 48 kHz stereo | | loadRegistry, resolveVoice, VoiceRegistry, RegistryVoice | The speaker-to-voice registry format |

Each provider takes its API key in its constructor.

Contracts

• Transcripts can be revised. A transcript event is keyed by utteranceId, and any field (text, speaker, even final) may arrive again for the same id. Treat the latest event per id as the truth. AssemblyAI revises text and speaker labels after an utterance is final; Scribe sends growing partials and then a commit.
• Speaker labels are optional. A per-speaker source, such as Discord's receiver streams, runs one session per speaker and leaves speaker unset. On a mixed feed, a diarizing provider fills speaker with its own opaque label, which can also be revised.
• PCM in, PCM out. Audio is PCM16LE mono. An STT session declares its input rate when it opens, and a TTS provider declares its output rate (outputRateHz). Converting to and from a surface's format is the consumer's job; audio.ts has the helpers.
• TTS timing is optional. When a provider has character timing for a chunk, onAlignment fires before that chunk's onAudio, with times in milliseconds from the start of the stream and characters matching the text as it was sent. A provider without timing never fires it.

Development

npm install
npm run build      # tsc → dist/
npm test           # bun test

Releases publish from a v* tag through GitHub Actions with npm trusted publishing (see .github/workflows/publish.yml).

License

MIT