Bland AI Uses a Second Model to Teach Speech AI to Laugh on Cue

October 8, 2026, 20:38Source: Bland AI

Bland AI says its research team developed NVAlign, a training method that uses a second model to listen to generated speech and check whether tags such as [laughs] or [sighs] were performed correctly. In a listening test on its English production system, accuracy rose from 39.8% to 54.0%, according to the company.

Key points

Bland AI's NVAlign uses a fixed speech-recognition model to score whether generated speech performs tags like [laughs].
On the English production system, accuracy in a 100-text listening test rose from 39.8% to 54.0%, according to Bland.
Naturalness ratings moved from 3.89 to 3.95 out of five.
English word-error rate increased on both evaluation sets, and one automated perceptual measure declined.
The method covers discrete tagged sounds and does not decide when a laugh is appropriate.

Why a simple tag is hard to perform

A developer can add a tag such as [laughs] or [sighs] to text in seconds. For a speech model, following that instruction is far harder. Bland uses the line "I thought you were serious. [laughs] You almost had me." as an example. The model must place the laugh at the right moment, keep the same voice, return to speech naturally and still pronounce every word correctly.

According to Bland, training on recordings paired with transcripts that label non-verbal sounds was not enough. Clean, labeled examples of less common sounds are hard to find, and even laughter varies widely between speakers. After that training, the models could produce many of the sounds but still skipped or confused them.

How NVAlign works

Bland's research team trained a separate speech-recognition model to recognize tags and non-verbal sounds. That model is then kept fixed and used to grade the output of the speech model. During NVAlign training, the process runs in four steps:

The speech model receives text containing a tag such as [laughs].
It generates the audio.
The recognition model scores how confidently it can detect the requested sound.
That feedback is used to adjust how the speech model plans and produces the audio.

Passing feedback through every step of audio generation would need a large amount of memory and computation. The team developed a shortcut that uses two selected steps to calculate the training update. Bland says developers do not need to manage this process when generating speech, because it happens during model training. The method is described in the NVAlign paper, with Qiaolin Wang named among the authors.

The company also describes a risk: a model can raise its tag-following score while making the rest of the audio worse, for example by forcing the sound too loudly, changing the speaker's voice or distorting the surrounding sentence. Bland saw this when it removed safeguards during testing. It added penalties for changes that damage speaker identity or audio quality, produce excessive peaks or low volume, or move too far from the model's existing behavior.

What the tests showed

Bland tested NVAlign on two public speech models and an English production system. It also ran a separate listening test using 100 texts. Three listeners judged, without knowing which system produced the audio, whether each requested sound appeared in the correct position. A result counted as correct only when all three agreed.

For the English production system, accuracy on that test rose from 39.8% to 54.0%. Listeners rated naturalness at 3.95 out of five, compared with 3.89 before NVAlign training.

There were trade-offs. English word-error rate increased on both evaluation sets, and one automated measure of the vocalization's perceptual effect declined. Bland says it still needs to check pronunciation and how each sound fits the sentence, even when the model follows the tag correctly.

What it means for customers and what is still open

Bland says the benefit for developers is more dependable control over delivery through the text they supply, so a requested laugh or sigh is more likely to appear where they placed it. The company lists several implications for customers:

Developers can direct delivery inside the text instead of editing audio after generation.
Voice agents can respond with more than words while staying in the flow of a call.
A single voice can carry changes in pacing, emphasis and emotion across a conversation.
Teams can design more specific moments without rebuilding the surrounding voice experience.

Bland is clear about the limits. The work focuses on discrete, tagged sounds. It does not establish that a model knows when laughter is appropriate or can handle every continuous change in emotion and speaking style, and the developer still chooses where to request a sound. Rare sounds remain harder to train because good examples are scarce, and both the speech model and the recognition model need that variety to improve.