Audio Coach: how new voices sound and what they cost

Ticket #774 · 13 October 2026 · English, Spanish and Ukrainian, coached in German

The Audio Coach is the personal recording each learner gets, about 25 minutes long. It has two parts. In the Vocabulary Trainer the coach drills the learner's hardest words, and in Passive Integration it reads example sentences with their translations. Today a voice from OpenAI reads it. We had the same Audio Coach read by four other voices to compare how they sound, how reliable they are and what they cost. In each language every voice reads exactly the same script, with the same pauses as today.

In short

Google's Gemini Flash Lite is the one to beat. It didn't leave out a single word in 869 lines, it speaks at today's pace, and it costs half as much as today's voice. Its Ukrainian came through word for word, where both OpenAI voices mispronounce it. OpenAI's basic voice is the cheapest, but it leaves out single words and rushes, and slowing it down makes it sound worse. Reusing parts that are already recorded can cut the cost roughly in half again, see "Paying for less" below.

Paying for less: reuse what's already recorded

Problems with OpenAI's basic voice

Listen

Each player plays one full Audio Coach in that voice. Under "Same moment", every button plays the same short passage in a different voice, so you can compare them back to back. Red buttons jump to words that OpenAI's basic voice left out.

More Gemini Flash Lite voices

All recordings side by side

How it sounds: a blind test

How we made and checked the recordings

Same script for everyone. The scripts come from the same tool that writes real Audio Coaches. The Spanish word list is from a real Audio Coach. The English and Ukrainian lists are made up but look the same.
Same delivery as today. Today's coach gets a written instruction to speak slowly and clearly, like a friendly teacher. The Gemini voices got the same instruction. OpenAI's basic voice can't follow instructions, so we also made a slowed-down copy of it to match today's pace.
Same sound as today. Every line is recorded on its own and joined with today's pauses, at today's sound quality.
Checks. Every recording plays to the end and is exactly as long as expected. A program measured every line to find the silent ones, and automatic transcription compared what was said with the script.

Full scripts

Every line has a play button for each voice. Red lines are ones that OpenAI's basic voice left out.