Recipes

Full offline turn with real engines

import { VoicePipeline, WhisperCppStt, OllamaLlm, PiperTts } from 'whispa';

const pipeline = new VoicePipeline({
  stt: new WhisperCppStt({ model: 'models/ggml-base.en.bin' }),
  llm: new OllamaLlm({ model: 'llama3.2' }),
  tts: new PiperTts({ model: 'voices/en_US-amy-medium.onnx' }),
  systemPrompt: 'You are a concise voice assistant.',
});

const { transcript, reply, audio } = await pipeline.turn('recording.wav');

Text-only assistant (skip STT/TTS)

import { VoicePipeline, MockStt, MockTts, OllamaLlm } from 'whispa';

const chat = new VoicePipeline({ stt: new MockStt(''), tts: new MockTts(), llm: new OllamaLlm({ model: 'llama3.2' }) });
const reply = await chat.respond('Summarize today in one line.');

Bring your own LLM engine

import type { LlmEngine, Message } from 'whispa';

class OpenAILlm implements LlmEngine {
  async chat(messages: Message[]): Promise<string> {
    const res = await client.chat.completions.create({ model: 'gpt-4o-mini', messages });
    return res.choices[0].message.content ?? '';
  }
}

const pipeline = new VoicePipeline({ stt, llm: new OpenAILlm(), tts });

Push-to-talk loop

async function onButtonPress() {
  const { reply, audio } = await pipeline.turn('capture.wav'); // your recorder writes this
  await playback(audio);
}

Cap the conversation memory

new VoicePipeline({ stt, llm, tts, systemPrompt: '…', maxHistory: 12 }); // keeps last 12 turns