Receitas

Turno offline completo com motores reais

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: 'És um assistente de voz conciso.',
});

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

Assistente só de texto (saltar 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('Resume o dia de hoje numa linha.');

Usar o teu próprio motor de LLM

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 });

Loop push-to-talk

async function aoPremirBotao() {
  const { reply, audio } = await pipeline.turn('capture.wav'); // o teu gravador escreve isto
  await reproduzir(audio);
}

Limitar a memória da conversa

new VoicePipeline({ stt, llm, tts, systemPrompt: '…', maxHistory: 12 }); // mantém os últimos 12 turnos