AgentTested hands-on review
tool: Coqui XTTS v2
slug: coqui-xtts
date: 2026-10-21
score: 6/10

Coverage: Built-in voice plus one clone. Intelligibility from whisper.cpp. No listening test.
What we tested: The site's TTS script with a built-in voice, plus voice cloning from a 12.8 s reference clip (our Kokoro output). We scored intelligibility by transcribing each clip with whisper.cpp.
Method: coqui-tts 0.28.0 (the maintained idiap fork) with torch 2.14 CPU, tts_models/multilingual/multi-dataset/xtts_v2, COQUI_TOS_AGREED=1. The first run crashed until we added the `codec` extra (torchcodec). WER from whisper.cpp base.en (_battery/wer.sh).

Notes:
- CPU-only box: 8 cores, 15 GB RAM (about 4-10 GB free during tests), no GPU.
- No listening or speaker-similarity test
- Clone run shared the CPU with an npm install
- Weights are Coqui Public Model License, non-commercial.

Results:
download_and_load_s: 29.8
builtin_voice: 140.1 s for 15.2 s of audio (9.2x slower than real time). WER 0.030 ('AgentTested' heard as 'Agentested')
voice_clone: 191.9 s for 13.2 s (14.6x slower; an npm install was running at the same time). WER 0.061
peak_rss: about 4.5 GB
listening_test: none

Pros:
- Very intelligible: WER 0.03-0.06, as good as Piper and Kokoro
- Clones a voice from one short clip, with no training
- Supports many languages

Cons:
- About 9-15x slower than real time on CPU
- Needed an extra package fix (torchcodec) before it ran
- Model licence (CPML) is non-commercial only
- Uses about 4.5 GB of RAM

Verdict: XTTS v2 spoke our script almost perfectly and cloned a voice from a 13-second clip without any training. On a CPU it is slow, taking more than two minutes for 15 seconds of speech, and the first run crashed until we added a missing audio package. The model licence forbids commercial use. We didn't listen to the clips or measure how close the clone sounds, only how clearly it speaks.

Score line: Accurate speech and easy voice cloning, but slow on CPU and non-commercial.