Modulate Built Voice Tools For Gamers. Now It Grades AI Agents.

On September 28, Boston startup Modulate announced $25 million in new funding led by Future Ventures, with Hyperplane and Lakestar joining. According to PitchBook data cited by TechCrunch, the company had raised $41 million at a $170 million valuation before this round.

Today Modulate sells enterprises a set of voice models that transcribe calls, read emotion, detect deepfakes and AI generated music, and check whether AI voice agents follow the rules in regulated industries.

That is not where it started. Modulate was founded in 2017 by Mike Pappas and Carter Huffman, who met as physics undergraduates at MIT. Its first product was voice modulation for gaming. Then it built a tool to moderate toxic voice chat. Now it analyzes calls between people and AI agents.

Three markets in nine years sounds like a company that kept changing its mind. Look closer and it is a company that kept one core asset and kept finding bigger buyers for it. That difference matters for every founder thinking about their next move.

Here is the rough timeline, pieced together from coverage over the years.

Chapter one: voice for games. Modulate's early focus was changing how players sound in online games.

Chapter two: moderation. By 2022, the company's main product was ToxMod, a system that listens to game voice chat and flags harmful behavior for human moderators. That August, GamesBeat reported a $30 million round led by Lakestar. Customers at the time included Rec Room and PokerStars VR. The company said its models had been trained on more than 10 million hours of audio, and it had hit its milestones with just 27 people.

In August 2023, Activision announced it would bring ToxMod to Call of Duty, with a beta starting August 30 and a full launch alongside Modern Warfare 3 on November 10. That is one of the biggest online games in the world choosing a startup's model for a very public problem.

Chapter three: enterprise voice intelligence. The new round funds a platform of more than 100 models aimed at call centers, regulated businesses, and companies deploying AI voice agents.

The common thread across all three chapters is a deep understanding of how people actually sound in live conversation, not just what words they say.

In the 2022 GamesBeat interview, Pappas made a point that explains the whole story. He said voice moderation cannot rely on commodity transcription, because converting speech to text loses tone, volume, and whether someone is shouting. He also said no other company had built a data set focused on real social voice chat online.

That is the moat. Game voice chat is noisy, emotional, full of slang, and often hostile. A model that can tell friendly trash talk from real harassment has learned something hard about human speech.

Fast forward to 2026. Huffman told TechCrunch that a lot of companies do transcription, but few capture the full nuance of a conversation. The same skill that separated banter from abuse in a game lobby now separates a satisfied caller from an unhappy one, or a real voice from a cloned one.

The buyer changed. The core capability did not. That is why the move looks natural in hindsight.

One quote from Huffman is worth pinning above the desk of anyone shipping an AI agent.

He said many companies assume that a neutral or positive customer means a successful call, and a negative customer means a failed one. In practice, he said, people are often polite even to bots. They do not sound angry, but they leave very dissatisfied.

This is a trap for any team that measures AI quality with simple sentiment scores. Your dashboard can show calm, courteous conversations while customers quietly give up and churn.

If you run an AI support agent, a voice product, or even a chat assistant, try these checks:

• Measure outcomes, not tone. Did the customer's problem actually get solved? Did they contact you again within a week about the same issue?
• Track silent exits. Look at people who end a session politely and never come back.
• Read or listen to a sample by hand every week. A human reviewer will notice resigned politeness that a score misses.
• Compare against a human baseline. Run the same kind of request through a human team and compare repeat contact rates.

Modulate built a business on this gap. You can at least make sure it is not hiding inside your own product.

Modulate's technical choice is also worth studying. Instead of one giant model, it runs more than 100 smaller models. TechCrunch described two groups: signal extraction models that read emotion, tone, language, and whether a voice is synthetic, and analysis models that judge intent, rule violations, and scam attempts. An orchestrator calls the ones it needs.

Huffman said this approach avoids the need for specialized hardware and heavy compute, which matters as token bills rise. It also makes it easier to train a new capability and add it to the set.

For founders, the lesson is about cost structure and speed. If your product depends on calling a large general model for every task, your margins move with someone else's pricing. Narrow models built on your own data can be cheaper to run and harder to copy. The company is also working on on-premises and on-device deployment, which opens doors with buyers who care about privacy.

Most startups do not need a dramatic pivot. They need to notice when their best asset has a bigger buyer. Here is a simple way to look for one.

1. Name your asset in one sentence. Not your product. The thing underneath it. For Modulate, it was understanding tone and intent in messy live speech. For you, it might be a data set, a workflow you automate better than anyone, or trust with a hard to reach buyer.
2. List who else has the same pain. Modulate went from game studios to call centers to enterprises running AI agents. All of them need to know what is really happening in voice conversations.
3. Check who pays more and buys faster. Gaming moderation is valuable, but regulated industries facing deepfake fraud and compliance rules have larger budgets and sharper deadlines.
4. Keep the old customers while you test. Modulate did not abandon gaming overnight. It kept shipping for existing buyers while it built for new ones.
5. Write down what would prove the move is working. Pick two or three signals, such as paid pilots or repeat purchases in the new segment, and a date to review them.

A written plan helps you avoid drifting. Some founders sketch this in a planning tool like Foundra so the new segment, the old segment, and the runway all sit on one page. Any clear document will do. The point is to make the move on purpose.

Adjacent moves are not free.

Focus. A 40 to 45 person team serving gaming, call centers, and AI agent compliance at once can get stretched. Modulate says it plans to add about 10 people, mostly to build models. Hiring to match a wider scope is a real cost.

Crowded categories. TechCrunch noted that Modulate competes with companies analyzing conversation intent and others fighting deepfake calls. The bigger the market, the more rivals arrive.

Proof in a new market. A strong gaming track record does not automatically convince a bank or insurer. Expect to earn trust again with pilots, security reviews, and references in the new segment.

Brand confusion. If buyers in one market know you for something else, you may need a clear new product name or story so each audience understands what you do for them.

How much did Modulate raise in September 2026? Modulate raised $25 million led by Future Ventures, with participation from Hyperplane and Lakestar, according to TechCrunch. PitchBook data showed $41 million raised at a $170 million valuation before this round.

What does Modulate do now? It runs more than 100 small voice models that handle transcription, emotion analysis, deepfake and AI music detection, scam detection, and policy checks for AI voice agents, mostly for enterprises and call centers.

How did Modulate start? It was founded in 2017 by Mike Pappas and Carter Huffman and began with voice modulation for games. It then built ToxMod, a voice chat moderation tool that Activision brought to Call of Duty in 2023.

What is the main lesson for first-time founders? Identify the asset underneath your product, then look for buyers with the same pain and bigger budgets. Moving markets works best when the core capability carries over.

Why does "polite but dissatisfied" matter for AI products? Simple sentiment scores can make AI conversations look successful when customers are quietly giving up. Measure whether problems get solved and whether customers return, not just how calm they sound.

1. Modulate raises $25M for its voice models and analysis suite, TechCrunch, September 28, 2026
2. Modulate gets $30M to detox game voice chat with AI, GamesBeat, August 17, 2022
3. Call of Duty partners with Modulate to add real-time voice chat moderation to Modern Warfare 3, Windows Central, August 30, 2023
4. Introducing ToxMod, Modulate blog (archived)