Speko: Optimized Solution for Voice Models
Automatically select the best STT, LLM, and TTS models for voice assistants.
1Speko's Purpose
Speko aims to find the best combination of speech-to-text (STT), large language models (LLM), and text-to-speech (TTS) models for voice assistants. The platform selects the most suitable models based on user-defined criteria, improving performance. This allows users to be aware of better and cheaper alternatives.
2Industry Problem
Voice assistants typically work with models chosen once and never updated. Even when new models hit the market, integration challenges and performance comparisons prevent changes. This results in users working with outdated and less efficient models.
3Speko's Solution
Speko selects the most suitable models based on user optimization criteria (accuracy, latency, cost, or balanced) and language/region information. The platform compares models, identifies the best one, and provides detailed information to the user. This process ensures voice assistants remain up-to-date.
4Technical Details
Speko tests models in different dates and regions, making the results publicly available. The platform measures model performance in various scenarios, such as spontaneous speech, money, and date information. Additionally, an automatic scoring system for TTS naturalness was developed, producing results consistent with human choices.
5Customer Stories
One customer, unsure which model to choose, started routing all traffic through Speko. Another customer discovered their STT model had a high error rate and switched to better alternatives. A medical team found the STT model that best handles medical terminology.
6Open Source Gateway
Speko offers an open-source gateway for teams wanting to avoid an additional network hop in the audio path. This gateway, written in Go, operates as a single binary file and communicates over a local Unix socket. In BYOK mode, the gateway works completely independently.
7Cost and Growth
Speko's gateway and BYOK setup are offered for free. Charges apply for the hosted router and managed keys. Usage, which started in late June, has grown by an average of 25% weekly. The platform saw particularly high interest during launch weeks.
8Community Feedback
Speko aims to continuously improve by gathering feedback from the community. Questions include how users currently select voice models and how much they trust third-party benchmarks. This feedback contributes to further enhancing the platform.
Original article: Hacker News: Front Page — Speko: Ses Modelleri için Optimize Edilmiş Çözüm