AssemblyAI vs Hume AI

Converts speech into text wherever the user is typing.

Safety add-ons for transcription requests: profanity filtering, content moderation and redaction of personal data in the transcript text and in the audio itself, each toggled per request.

OpenAI-compatible API that routes requests to GPT, Claude, Gemini, Qwen and other hosted models with automatic fallbacks, at the providers' own token prices with no markup, beside AssemblyAI's speech APIs.

Speech-to-text API for recorded audio and video files: transcripts in 99 languages with speaker diarization, language detection, word-level timestamps and keyterm prompting, billed per second of audio.

Live transcription API: audio streamed over a WebSocket comes back as text within a few hundred milliseconds, with endpointing for voice agents, sold by the hour of session time and also offered self-hosted.

Add-on models switched on in a transcription request that turn recorded audio into structured output: speaker labels, sentiment, entities, topics, key phrases, summaries, translation and custom formatting, each priced per audio hour.

One connection that runs a production voice agent end to end: speech recognition, a conversational language model, voice generation, turn detection, interruptions and tool calls, at one flat hourly rate.

Converts spoken audio into text with a cleanup pass and an accompanying verbatim transcript.

Returns a transcript of a short audio clip in a synchronous API response.

Creator Studio is Hume's web application, hosted at app.hume.ai, for turning written documents into narrated audio: 'From document to audio, effortlessly. Import your content, assign voices, add acting directions, and export production-ready audio.' Hume targets it at creators producing audiobooks, videos, podcasts, documentaries, e-learning courses and corporate training that need multi-voice narration with emotional direction.

Empathic Voice Interface (EVI) is Hume's real-time speech-to-speech voice AI API, described in Hume's docs as 'an advanced, real-time emotionally intelligent voice AI.' The product page markets it as a way to 'Build voice-first AI experiences that understand and respond to human emotions' and states it 'works seamlessly with Claude, GPT, Gemini, Grok, Kimi K2, Llama, and more' - layering voice on top of third-party language models rather than supplying its own. It is billed per conversation minute on Hume's public plans.

The Expression Measurement API measures emotional and vocal expression in speech, marketed as 'Measure expression in voice, offline or in real time.' Hume describes two endpoints on the one page: 'The Tagger API returns 600+ expression and voice dimensions from any audio, at scale. The Prosody API delivers real-time emotional expressions during live conversations.' Measured dimensions include emotion categories, speaking styles, vocal qualities such as pace, warmth and hesitation, and expression intensity with confidence scores.

The Human Feedback API sells human evaluation of voice and conversational AI as an API call, marketed as 'One API call. Real human ratings.' and 'Plug human evaluation into your pipeline. Audio in, structured results out.' It supports three study types - live conversations with AI models, single-sample audio ratings with custom questions, and side-by-side A/B comparisons - with Hume handling participant recruitment, quality control and fraud detection. Hume states it powers 'evaluation for Hume's own frontier voice models.'

Octave TTS is Hume's text-to-speech API, which Hume's documentation calls 'the first text-to-speech system built on LLM intelligence.' The marketing page positions it as 'Text-to-speech with emotional intelligence' that generates 'expressive, natural-sounding speech that conveys the full range of human emotion,' with SDKs for Python, TypeScript, .NET and Swift and streaming audio output. It is billed by characters on Hume's public self-serve plans.

User Testing API runs tests of voice models with human participants.