Hume AI Pricing & Developer Hub

import { HumeClient } from 'hume';

const client = new HumeClient({ apiKey: process.env.HUME_API_KEY });

// Connect to Empathic Voice Interface with live prosody emotion telemetry
const socket = await client.empathicVoice.chat.connect({
  onMessage: (message) => {
    if (message.type === 'user_message') {
      console.log('Detected emotions:', message.models.prosody?.scores);
    } else if (message.type === 'audio_output') {
      playAudioBuffer(message.data);
    }
  }
});

Hume AI Empathic Voice Interface (EVI) Semantic Tone Modeling

Hume AI's Empathic Voice Interface (EVI) departs fundamentally from standard text-to-speech architectures by jointly modeling acoustic prosody alongside semantic token distributions. Traditional voice bots pipeline discrete STT, LLM, and TTS steps, losing critical paralinguistic nuances like pitch modulation, hesitations, breath cadence, and tonal inflection. Hume's multimodal transformer processes raw continuous PCM audio waveforms directly, extracting 48 distinct dimensions of emotional and expressive sentiment in real time. This bidirectional architecture enables instant conversational turn-taking, allowing users to interrupt the AI naturally without awkward multi-second latency buffers.

Hume WebSocket Audio Streaming & Buffer Congestion Management

Integrating Hume EVI into production web applications requires streaming 16 kHz or 24 kHz linear PCM audio chunks over low-latency WebSockets. The client application dispatches base64-encoded audio frames every 50 to 100 milliseconds. When network congestion degrades client bandwidth, Hume's edge ingress gateway buffers inbound frames and dynamically adjusts jitter-buffer sizing to prevent audio packet starvation. Client-side audio visualizers can bind directly to Hume's real-time emotion probability events to dynamically render reactive conversational avatars in the browser.