There is a persistent hope in the voice AI industry that emotional intelligence can be added at the end. Take a capable model, write a system prompt instructing it to be warm and empathetic, maybe route in a sentiment score, and ship. It is an appealing idea because it is cheap. We think it is wrong in a specific and instructive way, and the reasoning matters for anyone deciding where to invest in their voice stack.
What prompting can and cannot reach
Prompting shapes what a model says. It is far weaker at shaping what a model perceives and how it sounds, because those capabilities are fixed by training.
Start with perception. If a model's representations do not encode the acoustic signatures of hesitation, strain, or suppressed frustration, no instruction can make it hear them. Telling a model to "pay attention to the customer's tone" presumes the tone information survived into the model's internal state. A system trained primarily to map audio to words has been optimized to treat prosody as noise to see through, not signal to preserve. You cannot prompt your way into a perceptual capability the training never built. Perception is learned.
Expression has the same structure. A model's ability to produce contextually appropriate delivery, the right warmth for a reassurance, the right steadiness for bad news, is bounded by what its training taught it about the mapping between context and prosody. Prompt-level instructions like "sound empathetic" produce, at best, a generic empathy register applied uniformly to every situation. Real emotional appropriateness is situational and fine-grained, and models only acquire fine-grained behavior through fine-grained signal during training.
The industry has learned this lesson once already. Instruction-following in text models did not come from clever prompting of base models. It came from reinforcement learning against human feedback. Capabilities that involve judgment, appropriateness, and nuance get built into weights, or they do not exist.
The bottleneck is the reward signal
If emotional intelligence must be trained in, the immediate question is what supervises the training. Reinforcement learning is only as good as its reward, and this is where voice AI hits a genuine gap.
Task-oriented rewards are relatively easy to construct: did the agent resolve the issue, complete the booking, answer correctly. Emotional quality is different. The reward has to reflect human perception of the audio itself, per turn, across multiple dimensions at once. Was the delivery appropriate to the moment? Did the response acknowledge what the user's voice, not just their words, communicated? Did the system remain steady and consistent as the conversation evolved?
Three properties make this signal hard to build. It must operate on audio, not transcripts, because much of what it needs to score is not in the transcript. It must be anchored in large-scale human judgment with demonstrated reliability, because emotional appropriateness is defined by human perception, and there is no authoritative answer key to substitute for it. And it must be separable from task reward, so that optimization cannot trade one against the other. A model should not be able to buy task-completion score by steamrolling an upset customer, and it should not be able to buy warmth score by abandoning the task.
This kind of reward model is the scarcest asset in the voice AI training stack right now. It requires a scientifically grounded account of emotional expression, a large corpus of audio labeled against that account by diverse human raters, and the modeling work to turn those labels into a signal stable enough to train against. None of those pieces can be improvised quickly, which is precisely why so few exist.
The compounding loop
Once that signal exists, something important changes about how progress works. Evaluation and training stop being separate activities and become one loop. Evaluation identifies where a model falls short on the dimensions humans care about. Those same dimensions, expressed as a reward, drive the training that closes the gaps. The next evaluation measures the improvement and finds the next gap.
Labs running this loop can improve on emotional quality the way the field improved on accuracy: systematically, measurably, iteration over iteration. Labs relying on prompting improve until the prompt's ceiling, which arrives quickly, and then plateau while attributing the plateau to model limitations.
There is a strategic asymmetry hiding in this. Prompting is available to everyone, which means whatever it achieves is achieved by everyone, at the same time, for free. Training signal is not evenly distributed. The labs with access to rigorous, human-grounded emotional reward will compound an advantage that prompt engineering cannot replicate, because the advantage lives in the weights.
This is the thesis behind how we have built our stack at Hume: rigorous human-grounded evaluation on one side, and reward signal for reinforcement learning on the other, built on the same scientific foundation and the same labeled corpus. Measurement that ends in a report is a diagnosis. Measurement that feeds training is a treatment.
Voice AI's next competitive era will not be decided by which lab writes the best empathy prompt. It will be decided by which systems were trained, deliberately and measurably, to understand and respond like humans. That is a training-time outcome, and it starts with the signal.
If you are training voice models and want to close the loop between evaluation and improvement, we should talk.
