About · Finnish Pronunciation Practice

About this tool

Finnish pronunciation is notoriously precise — vowel length, consonant gradation, and stress patterns all carry meaning. Even advanced learners benefit from objective feedback on what they actually said versus what they intended to say.

This tool lets you type any Finnish phrase, speak it aloud, and immediately see a character-level diff between your intended phrase and what the speech recognition model transcribed. It's designed to help you identify specific sounds or patterns you're consistently missing, so you can focus your practice where it matters most.

What you can do

🎙️ Hold the record button to record yourself speaking the phrase — release to send.
📊 Each attempt is scored at the character level using CTC forced-alignment. A bar chart shows how confidently the model detected each individual sound, so you can see exactly which letters you nailed and which need work.
🎨 A "What I heard" column alongside the bar chart shows the model's full transcription with colour-coded highlights — green for matched sounds, red for missed or substituted letters — so you can see the overall difference at a glance.
🔊 Tap the speaker icon next to the phrase input to hear the expected Finnish pronunciation, or replay your own recording from each attempt row to compare.
📱 Works fully on mobile — no app install required. Designed for quick practice sessions anywhere.

How it works

Speech recognition is handled by Finnish-NLP wav2vec2-xlsr-300m, a fine-tuned variant of Facebook's wav2vec2 model trained specifically on Finnish speech. It combines acoustic features with a KenLM language model to produce accurate Finnish transcriptions.

The model runs on a dedicated GPU worker hosted on RunPod. When you submit a recording, your audio is sent securely over HTTPS to that worker, transcribed, and the text result is returned — typically within a couple of seconds. The worker may take some time to respond after a period of inactivity while it warms up.

Privacy & voice data

We take your voice data seriously. Here is exactly what happens when you record yourself:

🚫

Your recordings are never stored. Audio is sent directly to the RunPod inference worker over an encrypted HTTPS connection, transcribed in memory, and immediately discarded. No audio file is written to disk or retained by the server at any point.

🔒

No personal data is collected. We do not collect your name, email address, IP address, or any identifying information alongside recordings. The transcription result (text only) is returned to your browser and processed entirely client-side.

🖥️

Processed on a private GPU worker. Transcription happens inside an isolated RunPod worker — it is not sent to any third-party API (OpenAI, Google, etc.). The model runs entirely within our infrastructure.

We collect anonymous, aggregate usage statistics (page visits). No cookies are set and no personal data is stored or shared.

Built by

Semolina Solutions Oy