- Which Hume were you using?
- Emotion and sentiment from speech
- Testing voice agents and shaping how they sound
- What's shifted in this corner of voice AI
- An hour with your own recordings
- Mistakes that cost Hume switchers a rebuild
- What you'll pay for, in general terms
- Where I'd start, depending on what you're building
Choose a Hume AI alternative by the job you hired Hume for. For reading emotion from a voice, try audEERING, Vokaturi, Imentiv AI or Affectiva. For testing voice agents, look at Coval, Hamming and Cekura. AssemblyAI adds sentiment to transcripts, Deepgram covers several voice layers in one API, and ElevenLabs handles expressive speech.
Plenty of teams first came to Hume when they wanted software to hear how someone said something, not just what they said. Its homepage now tells a different story. It leads with measuring and improving voice AI performance, describes an evaluation platform for models and agents, and pitches a new Audio Decisions API as one model that answers any question straight from audio. So before you shortlist a replacement, work out which Hume you were using, because each version has a different kind of rival.
Which Hume were you using?
I'd sort it into four jobs before you look at a single vendor:
- Emotion and expression from the voice itself. You want a read on anger, calm or stress from tone and pace, whatever the words were. That's speech emotion recognition, a small and specialist field.
- Sentiment on what was said. Often a product team only needs to know whether the caller was unhappy, and the transcript can tell you that. Speech-to-text APIs now sell it as an add-on layer.
- Checking that a voice agent behaves. If you came to Hume through its newer evaluation pitch, your real alternatives are voice agent testing platforms that run simulated calls and score what happens.
- Making a voice sound like it means it. That's generation, not measurement, and it lives with the text-to-speech APIs rather than with emotion vendors.
The first two get mixed up most, and it isn't a harmless mix-up. A model that listens to tone can score a politely worded complaint as calm. A model that reads the transcript will catch the complaint and can miss sarcasm delivered in a flat voice. Neither is wrong; they're answering different questions.
Emotion and sentiment from speech
audEERING
A specialist audio analysis company whose devAIce toolkit gives developers expression recognition from voice, for teams building it into their own product.
Vokaturi
If five basic emotions cover your use case, this speech emotion software offers them off the shelf, with a section aimed at developers.
Imentiv AI
Reads emotion across video, audio, text and images rather than voice alone, with real-time analysis and an SDK that runs offline.
Affectiva
Emotion AI from Smart Eye for media analytics and in-cabin sensing, with an emotion SDK for product teams.
AssemblyAI
Developers who mostly need sentiment, summaries and entities on top of transcripts get that from this speech-to-text API's understanding layer.
Deepgram
A broad voice AI API platform covering speech-to-text, text-to-speech, voice agents and audio intelligence, so one vendor can handle several layers.
Testing voice agents and shaping how they sound
Coval
Built to simulate and evaluate voice agents before launch and after it, as a single evaluation layer for the team shipping them.
Hamming
Throws simulated calls at voice and chat agents to find breakages early, then keeps watching the live calls once you've shipped.
Cekura
Automated QA for voice and chat AI agents from a small team that also publishes public voice AI benchmarks.
ElevenLabs
On the generation side: text-to-speech models you can tune for consistency, latency or emotional control, plus a platform for building voice agents.
What's shifted in this corner of voice AI
Two things changed, and they explain why a list of Hume alternatives now mixes emotion vendors with testing tools.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
First, the thing being measured is often a machine. When an AI agent answers the phone, the question stops being how the customer felt and becomes whether your agent handled them well. Testing platforms grew into their own category to answer that, and three of the ten tools here do nothing else.
Second, emotion got folded into bigger APIs. Transcription vendors sell sentiment, summaries and entities as a layer on the transcript, so a plain positive-or-negative score no longer needs a specialist contract.
What hasn't changed is how coarse off-the-shelf voice emotion still is. Vokaturi, for example, returns five basic emotions out of the box: happy, sad, neutral, angry and scared. If your product promises to spot boredom, confidence or confusion, check exactly which labels a model outputs before you design a feature around a word it never returns.
An hour with your own recordings
Don't judge any of these on a demo clip. Demo audio is chosen to work. Pull a dozen real recordings from your own product instead.
Pick twelve clips: four where the speaker was clearly upset, four where they were clearly fine, and four awkward ones, such as a polite complaint, a joke said deadpan, or someone tired rather than cross. Label all twelve yourself before you run anything, so a tool's output can't nudge your judgement.
Run them through two kinds of tool: one that reads emotion from the voice and one that scores sentiment from the transcript. Count how often each agrees with your labels, but spend most of your time on the awkward four. That's where the two approaches split, and it's where you'll see which one matches the way your users talk.
If your job is agent testing, the hour looks different. Write three calls a real customer might make: one who interrupts mid-sentence, one with a strong accent, and one who asks for something the agent shouldn't do. Run each through a simulator and see whether the report tells you what broke and at which turn, or just hands you a pass rate. A pass rate with no transcript of the failure won't help you fix anything.
Mistakes that cost Hume switchers a rebuild
- Buying emotion when you needed sentiment. If the decision you'll make is about the conversation (escalate this call, flag this review), transcript sentiment from the API you already use may be enough, and it's one integration instead of two.
- Treating an emotion score as a fact about a person. It's a signal from a few seconds of audio. Use it to route, sort or prompt a human, never to make a decision about an individual on its own.
- Testing an agent only before launch. Agents drift as prompts, models and callers change. If customers talk to it, you want simulation before release and monitoring after it, ideally in one tool.
- Ignoring where the audio goes. Voice recordings are personal data. If that worries your legal team, an SDK that runs on the device or offline may matter more than a few points of accuracy. In the EU there's a disclosure duty too: under Article 50 of the AI Act, anyone deploying emotion recognition must inform the people exposed to it, so build that notice into your flow.
Do you make voice emotion, speech analysis or voice agent testing tools? Tool makers can ask to be considered for a featured place on this list.
What you'll pay for, in general terms
Nine of the ten don't show a price on the pages read for this guide, so expect to talk to sales for most of them. The one that does uses credits, which is common in voice generation. Rather than chase a number, find out what the bill is counted in, because that decides whether it grows with you or against you.
For emotion and sentiment APIs, ask whether you pay per minute of audio processed, per call, or through a flat licence for an SDK you ship inside your own app. For testing platforms, ask how simulated calls are counted, whether live-call monitoring is billed separately, and what happens when you run a big regression test the night before a release. For voice generation, check how many credits a minute of speech uses at the quality you'll ship, not the cheapest setting.
Where I'd start, depending on what you're building
- You're building a product that reacts to how users sound: a specialist speech emotion SDK. Check its label set and whether it can run offline.
- You're mining support calls for unhappy customers: a speech-to-text API with a sentiment layer, so the transcript and the score come from one place.
- You're shipping a voice agent: a testing platform that simulates calls before launch and monitors them after. Pair it with a voice agent platform you trust, not instead of one.
- You analyse video and text as well as voice: a multimodal emotion platform, so you aren't stitching three vendors together.
- You want speech that sounds expressive: a text-to-speech provider with emotional control. If you've already tried the obvious choice, the alternatives to ElevenLabs are worth a look too.
If you only remember one thing: name the decision your product will make from the audio, then pick the tool that answers that question. Hume's own move towards evaluation is a good reminder that the question shifts as your product grows.
Want your tool on this list? Featured spots and reviews here are paid partnerships. Here's how to partner with me.


