Conclusion First

The cost of implementing voice AI has “approached zero.”

The era of needing to contract cloud voice services costing tens of thousands of yen per month is coming to an end.

NVIDIA’s newly announced compact voice model, “Nemotron 3.5 Bolt E4B,” operates on just 4GB of VRAM. This means it can run on a second-hand GPU costing a few tens of thousands of yen or even an existing laptop. Block’s open-source voice model, “Dessa,” operates locally on a MacBook M4 Max and can engage in conversations that are indistinguishable from human dialogue.

Neither of these sends voice data to the cloud. They operate entirely locally. This has two significant implications for small and medium-sized enterprises: monthly API fees become zero and customer voice data does not leave the company.

The challenge is not to stop at “amazing technology.” So, which tasks should be tackled first?

What Are the Differences Between NVIDIA and Block’s Models?

First, let’s clarify the technology.

NVIDIA’s “Nemotron 3.5 Bolt E4B” is a compact model tuned with reinforcement learning based on Google’s “Gemma 4 E4B.” It operates on 4GB of VRAM and can be run on edge devices like NVIDIA’s RTX series or Jetson. It has recorded scores surpassing the base model across 15 benchmarks, achieving practical-level accuracy despite its small size. It is intended for use in conversational agents and real-time voice interactions.

The key point is the number “4GB.” It can run on an RTX 3060 (which can be found second-hand for 20,000 to 30,000 yen). There’s no need to purchase a new server.

Block’s “Dessa” takes a different approach. It is trained on actual human conversation data, capable of mimicking backchanneling, pauses, and hesitations. In demo videos, it creates the illusion of conversing with a human over the phone. It operates in a local environment on a MacBook M4 Max and is available as open-source.

The key point here is its “human-like” quality. This technology can be directly linked to automating phone responses and customer interactions, rather than just transcribing meeting minutes.

In summary:

Both models not only fall below “50,000 yen per month” but, with existing PCs, the additional monthly costs are nearly zero.

So, Which Tasks Should Be Automated First?

There are countless tasks that can be automated with voice AI. However, small and medium-sized enterprises should first tackle tasks that are large time sinks and not overly dependent on specific individuals.

Here are three candidates, along with estimated monthly costs and potential savings. The assumption is an employee hourly wage of 1,800 yen (including social insurance and indirect costs).

1. Phone Responses (Top Priority)

• Current Cost: 2 hours per day on average × 20 days per month = 40 hours/month → 72,000 yen/month
• After AI Implementation: Automate first contact (filtering sales calls, automatic responses to common inquiries) with Dessa. If calls that require human intervention are forwarded, response time can be reduced to less than half.
• Savings: 20 hours/month × 1,800 yen = 36,000 yen/month savings
• Payback Period: Even with a new MacBook M4 Max costing about 250,000 yen, the payback period is 7 months. If an existing Mac is available, profits can be realized from the first month.

Why is this the top priority? Phone responses have a physical constraint of “someone has to answer.” By eliminating this, the person’s original tasks—sales, planning, customer support—can regain valuable time. The savings are not just financial; it also recovers lost opportunities.

2. Meeting Minutes (Significant Impact but Caution Required)

• Current Cost: 60 minutes for one meeting + 60 minutes for minute-taking = 2 hours. 8 meetings a month = 16 hours → 28,800 yen/month
• After AI Implementation: Local transcription and summarization based on Nemotron can reduce minute-taking to 5 minutes.
• Savings: 8 hours/month × 1,800 yen = 14,400 yen/month savings
• Payback Period: If the model is installed on an existing PC, initial costs are almost zero. Profits can be realized from the first month.

However, caution is warranted. The automation of meeting minutes hinges on whether “80% accuracy is acceptable.” This is fine for internal meetings. However, if automating minutes for client meetings, human verification is essential. Initially focusing solely on internal meetings is more practical.

3. Daily Reports and Documentation (Unassuming but Certain)

• Current Cost: 30 minutes per day × 20 days = 10 hours → 18,000 yen/month (per person)
• After AI Implementation: Voice input via smartphone while on the move or after tasks → AI formats it into a standard template. Task time is reduced to under 5 minutes.
• Savings: 8 hours/month × 1,800 yen = 14,400 yen/month savings (per person)
• For 5 Sales Staff: Monthly savings of 72,000 yen.

Daily reports often fall into a negative loop of “being tedious to write → resulting in thin content → leading to reduced readership → becoming a formality.” By using voice input, the barrier to writing disappears, increasing the amount of information. Consequently, daily reports can transform from a “management tool” to a “knowledge-sharing tool.” While this may not be easily quantifiable, it holds significant value for the organization.

Summary of Priorities

The Rule is to “Start Small and Let It Spread Naturally”

A common mistake for small and medium-sized enterprises is to launch a “company-wide implementation project.” Requirements definition, vendor selection, and approvals can take six months.

The greatest advantage of local AI is that you can start testing today on a single PC.

Here’s how to get started:

1. Choose one task first. Whether it’s phone responses, daily reports, or meeting minutes. The criterion is “the task that consumes the most time and is the most tedious.”
2. Install the model on an existing PC. For Nemotron 3.5 Bolt E4B, use a PC with an RTX; for Dessa, use a Mac. Installation instructions are available on GitHub.
3. Use it alone for a week. Identify issues with accuracy and UI.
4. If it works well, pass it to the person next to you. No manual is needed. Just say, “Try this out.” Good tools spread naturally.

Initial costs are zero if using an existing PC. Even if a new GPU-equipped PC is needed, it will cost between 100,000 and 250,000 yen. Monthly cloud costs are zero. Annually, automating phone responses alone can save over 430,000 yen.

What Truly Changes is Not the “Cost”

While we’ve discussed costs up to this point, to be honest, saving 30,000 to 70,000 yen per month is not a life-changing amount for small and medium-sized enterprises.

What truly changes is how people are utilized.

Administrative staff who were answering phones can now focus on customer follow-ups. Salespeople who spent 30 minutes writing daily reports can make one more visit. Young staff who were creating meeting minutes can concentrate on the content of the meetings.

Voice AI does not take away jobs; it takes away the time spent pretending to work.

Large corporations invest tens of millions of yen to automate call centers with AI. Small and medium-sized enterprises can implement the same for under 50,000 yen, using their own hands. This asymmetry is precisely why this technology deserves attention now.

Start with one PC, one task, and try it out within this week.