Audio Coach with text-to-speech: four options compared

Ticket #774 · generated 13 October 2026 · English, Spanish and Ukrainian, coached in German · 12 recordings, about 5½ hours of audio

Each recording is a complete Audio Coach in today's production format: the Vocabulary Trainer (intro, study language → German, German → study language, outro), then the Passive Integration track (intro, sentence pairs, outro). The pauses and the MP3 encoding are production's. Within a language all four options read exactly the same script, written by the production script generator for 20 words and 5 grammar items. Production currently voices the Audio Coach with OpenAI gpt-4o-mini-tts.

In short

Weak points of OpenAI tts-1

Listen

Each player is one full Audio Coach. The buttons under "Same moment" play one short passage in each version, so the voices can be compared directly. Red chips jump to words that tts-1 left silent.

Numbers

Blind listening impression

How the recordings were made and checked

Script. One script per language from the production script generator (the same prompts and model). English and Ukrainian use made-up word lists in the production format; the Spanish list is a real Audio Coach list.
OpenAI tts-1: standard model (not HD), voice alloy like production, speed 1.0 like production. tts-1 cannot take style instructions.
OpenAI tts-1, slowed: the same request at speed 0.6, the value at which tts-1 matched production's measured pace (10.0 characters per second). Generated by calling the OpenAI API directly with the same script, and joined with the same pauses and encoding.
Gemini 3.8 Flash TTS and Flash Lite TTS: voice Kore. Gemini has no numeric speed, so it receives production's style text ("clear, slowly, articulate…") on exactly the segments where production sends it.
Format. Each segment is spoken separately and joined with production's pauses (1, 2 or 3 seconds), then encoded the way production encodes: MP3, 64 kb/s, 24 kHz, mono.
Checks. Every file decodes to the end without errors, and its length equals the sum of its spoken segments plus the pauses. Every segment's loudness was measured to find silent ones. Each file was transcribed twice, once in the study language and once in German, and compared with the script. No style text was read aloud in any file.

Scripts

Every line can be played in each model. Lines in red are silent or cut short in tts-1.