FlowSTT

FlowSTT

Voice products on open speech models

FlowSTT is a set of voice products: a diary for language learners, an interview trainer and a family memoir recorder. Speech recognition runs on self-hosted open models.

See the products How it works

Products

Access is by invitation for now. To ask for access, write to support@flowstt.com.

Voice Diary

Invite only

Talk about your day in the language you are learning. You get the transcript and a tutor's review: mistakes, a corrected version, more precise words and exercises built from your own errors.

Local speech models, cloud LLM API, Telegram

Interview Trainer

Invite only

The bot asks an interview question, you answer by voice and get feedback: a checklist, what to fix, a model answer and a follow-up question. Questions cover Go, databases, system design and behavioral topics, in English or Russian.

Local speech models, cloud LLM API, Telegram

Memoirs

Invite only

A parent dictates memories by voice. The bot transcribes them, asks gentle follow-up questions about missing periods and details, and assembles the recordings into a memoir written in the narrator's own voice.

Local speech models, cloud LLM API, Telegram

Voice Support Agent

Coming soon

A phone agent for customer support: a call comes in, speech is recognised as it is spoken, an LLM answers using tools such as order status, and the reply is spoken back. Built on LiveKit.

LiveKit, streaming speech recognition, LLM, speech synthesis

How it works

The products share one idea. Audio goes in, open models running on a CPU-only server turn it into text, and an LLM does the language work on that text.

Processing pipeline A voice message goes to speech recognition (Parakeet, GigaAM, Whisper: open models, self-hosted, no GPU). Speech recognition produces text. An LLM (cloud API) does the review, follow-up questions and memoir work on the text. The result is a reply in chat. Voice message Speech recognition Parakeet, GigaAM, Whisper. Open models, self-hosted, no GPU Text LLM Review, follow-up questions, memoir. Cloud API Reply in chat Processing pipeline A voice message goes to speech recognition (Parakeet, GigaAM, Whisper: open models, self-hosted, no GPU). Speech recognition produces text. An LLM (cloud API) does the review, follow-up questions and memoir work on the text. The result is a reply in chat. Voice message Speech recognition Parakeet, GigaAM, Whisper. Open models, self-hosted, no GPU Text LLM Review, follow-up questions, memoir. Cloud API Reply in chat

In Memoirs the same recording goes through three engines, and an LLM merges the three transcripts into one.

Speed and memory of the speech engines on the production server (2 vCPU, no GPU), measured in October 2026
Engine Real-time factor Peak memory
GigaAM v3 (int8) 0.09 to 0.11 0.6 to 0.7 GB
Parakeet TDT 0.6B v3 (int8) 0.14 to 0.17 1.6 to 2 GB
Whisper large-v3-turbo 1.62 2 to 3 GB

Real-time factor is processing time divided by audio length: 0.1 means ten seconds of audio take one second to process. Lower is faster. Whisper large-v3-turbo is slower than real time on two CPU cores, so Parakeet is the main engine in Memoirs.

Technical notes

Write-ups of the measurements and the design behind the products.

  • Three speech engines on two CPU cores: speed and memory In preparation
  • Merging transcripts from several engines with an LLM In preparation