AgentTested hands-on review
tool: Bark
slug: bark
date: 2026-10-21
score: 4/10

Coverage: TTS script plus a [laughs] line. Intelligibility from whisper.cpp. No listening test.
What we tested: The site's TTS script, plus a short line with a [laughs] cue. Intelligibility scored with whisper.cpp.
Method: transformers 5.20 BarkModel suno/bark-small, voice v2/en_speaker_6, torch CPU, seed 0.

Notes:
- CPU-only box: 8 cores, 15 GB RAM (about 4-10 GB free during tests), no GPU.
- Small model only
- No listening test

Results:
full_script: 383.3 s for 13.1 s of audio (29x slower than real time). WER 0.121: 'clean Linux box' came out as 'clean one Xbox'
laugh_line: 167.8 s for 5.7 s (29x). Laugh not verified by ear
peak_rss: about 2 GB

Pros:
- MIT licence, fully local
- Can do non-speech cues like laughter (not verified by ear)

Cons:
- Extremely slow on CPU: 6+ minutes for 13 seconds of speech
- Least accurate TTS we've tested (WER 0.12, garbled 'Linux box')
- Output is random from run to run

Verdict: Bark worked, but it was the slowest and least accurate voice tool we've run. It took over six minutes to speak 13 seconds and turned 'Linux box' into 'one Xbox.' Its strength is expressive sounds like laughter, which we couldn't check by ear. For plain narration, Kokoro or Piper are far better on a CPU.

Score line: Works locally, but very slow on CPU and mangled a key phrase.