Quality review

Speech quality is reviewed before it becomes a public claim.

The benchmark page describes what RESONYX measures and what still needs human listening evidence. Public scores stay private until the sample set is large enough to be meaningful.

Review statusTargets, not claims

Current benchmarks describe dimensions, review rules, and sample gates. They do not publish comparative scores yet.

Review dimensions

What the listening panel checks.

Curated samples active

Indian English authenticity

Measures pronunciation, rhythm, phrase endings, and Indian professional speech cadence.

Human panel before public claim

Covered by Indian English, Hinglish, and regional-language demo samples with explicit reject rules for generic English fallback.

QA gated

Whisper authenticity

Checks airflow texture, consonant softness, breath placement, and intimate pacing.

Reject low-volume fake whisper

Whisper samples must show air, breath texture, softer consonants, and slower near-field pacing before being promoted.

Script-level controlled

Pause realism

Evaluates micro pauses, reflective pauses, sentence timing, and non-robotic comma handling.

Listening proof required

Launch scripts encode purposeful pacing; human listening review decides whether audio actually proves it.

Tracked as risk

Listener fatigue

Lower is better. Captures harshness, repetition, over-compression, and synthetic strain.

Lower fatigue wins

Samples that sound sharp, repetitive, compressed, or tiring are blocked even if they pass technical generation checks.

Methodology

How a sample becomes public evidence.

Curated sample set

Only approved public demo combinations are benchmarked. Unreviewed output is not treated as product evidence.

Blind A/B listening

System baseline and RESONYX realism must be compared without exposing implementation details to reviewers.

Reject rules first

A sample fails if whisper is only low volume, Hinglish sounds translated, or Indian English collapses into generic western cadence.

Score after proof

Public numeric scores stay private until enough human reviews exist. Public pages show evidence status and release gates instead.

Evaluation stack

Every generation path is reviewed with the same speech-quality lens.

NaturalnessBreath realismWhisper authenticityEmotional believabilityPause realismVocal texturePronunciation clarityAccent authenticityListener fatigueAI-detection risk

Release rules

What stays off the public page until evidence exists.

No comparative quality claim without human listening evidence.

No benchmark score is published until sample count and methodology are disclosed.

Every public demo must map to an approved sample manifest entry.

Regional-language demos are curated proof samples, not a broad translation guarantee.