LLM Tools|Index 04

Speko automates voice agent model optimization

Speko offers an API to continuously benchmark and switch between STT, LLM, and TTS models for optimal voice agent performance.

Via: AITECH TOKYO Editors
Dateline: Tokyo, August 17, 2026
Date: August 17, 2026
Time: 6 min read

Source

Hacker News Top

Speko automates voice agent model optimization

Tagline

An API to optimize voice AI models

Who & Why

For Tokyo-based product managers building enterprise voice agents, Speko automates the continuous selection and optimization of STT, LLM, and TTS models, ensuring optimal performance and cost efficiency across languages.

vs. Existing

Speko competes with the costly, manual process of benchmarking and integrating various STT, LLM, and TTS vendors, offering an automated API-driven approach that continuously optimizes performance without re-integration.

Tokyo Take

Speko offers a compelling solution for optimizing multi-language voice agents, crucial for Tokyo's global businesses. Its API-first approach is accessible, but true value for Japanese enterprises hinges on its specialized Japanese STT/TTS model benchmarking and localized pricing.

Speko is an API platform designed to continuously optimize the combination of speech-to-text (STT), large language model (LLM), and text-to-speech (TTS) components within production voice agents.

Founded by Bek, who previously built enterprise voice agents across Asia, Speko addresses the challenge of constantly evolving AI models. Companies typically evaluate models once and rarely update their stack due to the integration overhead and internal friction.

The platform functions as a router and gateway. Users send requests with optimization criteria—accuracy, latency, cost, or a balanced approach—along with language and region. Speko then selects the best-performing models from its public benchmarks.

This process replaces manual, labor-intensive benchmarking. Before Speko, teams would hire native speakers to rate new models, a ritual that Speko now automates via an API.

"we can literally go to this dashboard, switch the model, and it will do it for us."

The system offers failover during connection setup and pre-fetches session plans to minimize latency.

Speko maintains impartiality by not training or selling models itself. It transparently publishes benchmark results, even when its selections underperform alternatives. It tests spontaneous speech, long takes, and specific vocabulary like money and dates, recognizing that production demands differ from short demo clips.

An open-source gateway is also available, allowing teams to avoid extra network hops and keep API keys local. This BYOK (Bring Your Own Key) setup is free, with charges applying only to the hosted router and managed keys. Launched in late June, the platform has seen consistent weekly growth.

For a Tokyo-based professional managing a voice agent project, Speko offers a way to ensure their service uses the most current and cost-effective voice AI components without the continuous R&D burden. This could lead to better customer experience and operational efficiency, especially for multilingual deployments.

The Tokyo Editor's Read

What this AI story could mean for Tokyo in the years ahead.

Imagine you're building a voice assistant, like the kind you might use for customer service or an automated booking system. You need three things: something to turn speech into text (STT), a brain to understand and respond (LLM), and something to turn text back into natural-sounding speech (TTS). The problem is, there are dozens of companies offering each of these, and new, better ones appear every month. Speko acts like an independent expert who constantly tests all these components, figures out which combination works best for your specific needs (say, accuracy in Japanese, or low cost), and then lets your system automatically switch to that best option through a simple API. It's like having a tireless, impartial consultant for your AI voice components.

For Tokyo readers, this technology could significantly improve automated customer support lines, public information services, and even language learning applications. Imagine calling a bank or a government office and getting a truly natural, accurate voice interaction in Japanese, rather than the stilted, sometimes error-prone responses common today. It could also make it easier for businesses to offer high-quality multilingual support without hiring more staff, enhancing services for inbound tourists or international residents. The core impact would be on the quality and cost-efficiency of voice-based digital services.

Speko's API is available globally today, meaning any Tokyo-based developer or company can integrate it immediately. However, its true impact will likely be felt within the next 12-24 months. This timeframe accounts for the need for Speko to deepen its benchmarking specifically for specialized Japanese STT/TTS models and for Japanese enterprises to integrate such an orchestration layer into their existing, often complex, IT infrastructure. The adoption rate will hinge on how effectively Speko can demonstrate superior performance with Japanese language tasks compared to existing in-house solutions or single-vendor stacks.

While there isn't a direct Japanese counterpart offering an identical multi-vendor orchestration and benchmarking API for voice components, major players like NTT and SoftBank are deeply invested in developing their own STT, LLM, and TTS capabilities, often bundled as proprietary solutions. Companies like LINE also provide robust voice AI services within their ecosystems. However, these are typically vertically integrated. Speko's value proposition lies in its neutrality and optimization across vendors. The closest adjacent player might be a company like Notta, which focuses on high-accuracy transcription but doesn't offer the multi-model orchestration Speko does. The gap for a truly independent, multi-vendor AI voice component optimizer in Japan remains.

Editorial: AITECH TOKYO Editors

Adjacent Tools

LLM Tools

Amazon's AI Data Strategy Raises Concerns Over Rare Text Preservation

The e-commerce giant reportedly converts unique physical texts into digital training data, sparking debate on ethical data acquisition and cultural heritage.

Via AITECH TOKYO Editors · 6 min read

Source:TechCrunch AI

LLM Tools

Grok's Image Generation Raises Content Safety Concerns

An alleged misuse of X's Grok AI to generate explicit imagery from a childhood photo underscores the critical challenges in AI content moderation and platform responsibility.

Via AITECH TOKYO Editors · 6 min read

Source:TechCrunch AI

LLM Tools

AI's Mathematical Ceiling: Beyond Pattern Matching

A recent analysis highlights that even advanced AI models struggle with genuine mathematical reasoning, suggesting a fundamental gap in their ability to "think" like mathematicians.

Via AITECH TOKYO Editors · 5 min read

Source:Hacker News Top

← Back to grid

Speko automates voice agent model optimization

World AI tech, read from Tokyo. Once a week, in Japanese.

Adjacent Tools

Amazon's AI Data Strategy Raises Concerns Over Rare Text Preservation

Grok's Image Generation Raises Content Safety Concerns

AI's Mathematical Ceiling: Beyond Pattern Matching