AI speech enhancer comparison, October 2026: per-clip scores for seven speech enhancement tools on 59 real recordings

Per-clip scores from a comparison of seven AI speech enhancement runs (Diffio 4.6 Pro, Diffio 4.6 Flash, Adobe Podcast Enhance Speech v2, Cleanvoice, Auphonic Speech Isolation, ElevenLabs Voice Isolator and Descript Studio Sound) and the untouched originals, on 59 real English speech recordings (about 29 minutes): archived sermons, echoey rooms, phone and remote calls, wind and traffic, crowds and chatter, hum and machines, narration over music, and historic recordings. Run on 2026-10-22. Metrics: SCOREQ (no-reference speech quality; reported as gain over the original), word error rate (faster-whisper small.en on each output against an MAI-Transcribe-2 transcript of the original) and ECAPA speaker similarity to the original recording. Files: per_clip_scores.csv (59 clips x 8 versions, with each clip's license, credit and source URL), summary_overall.csv and summary_by_scenario.csv (means with 95% intervals), head-to-head counts, tools_and_settings.csv and metadata.json. No audio is included; every clip plays before and after, with its credit, on the write-up: AI speech enhancer comparison (Diffio AI). Disclosure: Diffio AI ran this test and makes two of the seven tools tested, so it is not an independent study. Every score is model-predicted; there was no human listening study. See README.md for methods and limitations. Diffio is not affiliated with or endorsed by Adobe, Descript, Auphonic, ElevenLabs or Cleanvoice.