RX 11 vs Adobe Podcast: Masking vs Regen, 12 dB, $/min

```html

The economics flip fast. Above roughly 12 dB SNR, regeneration wins the dollars-per-minute war outright: free, browser-based, near-instant. Below that line, spectral masking still owns the sub-12 dB sliver where regenerators smear timbre, plus the reverb, bleed, and confidentiality jobs no cloud upload should touch. RX's sticker price tops the $150-$400 repair band for exactly that reason — and no other.

Listeners changed first. British Council 2026 data puts 80% of English-language interactions between non-native speakers, and the emerging Intelligibility Standard grades clarity over accent. Corporate narration — the bread-and-butter volume business of voiceover — lives on that clarity. Paid suites span entry-level subscriptions to pro-grade perpetual licenses; outsourced human rescue runs on 48-hour turnarounds. Measure input SNR before you spend a dollar.

Dialogue Isolate is a discriminative neural source-separator: per iZotope's RX 11 documentation, it estimates a soft time-frequency mask over the STFT spectrogram, attenuating everything that is not speech while preserving the original waveform's phase and microphone character. It ships inside the RX 11 Standard suite as a standalone app and AU/VST3/AAX plugins, processing entirely on your local machine — so its worst case is bounded at residual noise. As Wikipedia's intelligibility entry notes, background noise and excess reverberation are what erode perceived clarity, conventionally expressed as signal-to-noise ratio; a mask removes the noise term without ever touching the speaker's identity.

Masking vs Regeneration

Enhance Speech v2, launched by Adobe, starts from the opposite premise. It is a generative cloud model that conditions a neural vocoder on the uploaded recording and regenerates a clean waveform rather than masking noise out of the existing one — hence the signature result: hyper-clean yet subtly re-spoken, with breath placement and micro-timing inferred rather than preserved. All inference runs on Adobe's servers, never your machine, which by itself disqualifies NDA-bound narration from that pipeline.

The control surfaces mirror the architectures. Adobe exposes a single 0–100% strength slider blending dry against enhanced signal — one degree of freedom, and easing it back is your only lever against timbre drift. RX exposes separation amount in decibels plus, new in RX 11, an integrated de-reverb stage inside the same module, clearing noise-plus-room-tone problems in one pass. As the essay "Why Dialogue Intelligibility is Important" argues, intelligibility is a stack of creative and technical decisions rather than a single setting; a dB-scaled control lets you match attenuation to the measured problem instead of trusting the model's judgment wholesale.

Here is the asymmetry that drives everything downstream. Mask-based separation cannot invent content, so its floor is residual noise — audible, measurable, survivable in review. Regeneration can hallucinate plausible-but-wrong phonemes and drift timbre, so its floor is identity loss — a voice that no longer verifies as your narrator's. Above the crossover threshold quantified in the next section, there is little for a generative model to invent and Adobe's gamble pays; below it, the model must supply speech the recording never contained, and hallucination risk scales with that deficit.

Throughput diverges structurally, too. RX runs at roughly real-time-or-faster on Apple Silicon with zero upload time; Adobe's wall-clock latency is dominated by upload plus cloud queue, neither of which you control at peak load. That reshapes effective cost per minute on deadline jobs independently of sticker price — and the fallback stings: according to the Medium essay "How AI Gave My Book a Voice," replacing narration outright runs $150–$400 per finished hour and often takes weeks, a bill local processing never forces.

Audition like a signal engineer, not a listener. Render one thirty-second clip through both engines, then phase-invert the dry file against each result: the RX output partially nulls — shared phase proving the mask kept the waveform's skeleton intact — while Adobe's refuses to null at all, because it is a new waveform. Then hunt the tells: masks leave breaths and plosive transients in place with hiss behind them; regeneration smooths or relocates breaths and rounds off consonant attacks. If the re-spoken quality reads on your narrator's voice, no per-minute discount buys it back.

Pricing settles what you pay; instruments settle whether it worked. POLQA (ITU-T P.863) is the naturalness gauge — full-reference, with a MOS that tops out at 4.5, so treat any vendor claiming higher as a red flag. MUSHRA (ITU-R BS.1534) runs a 0–100 scale against a hidden reference and answers the suppression-depth question: how much noise survives, and how ugly are the artifacts left behind. Google's ViSQOL v3 predicts wideband MOS without recruiting a panel — the budget stand-in. None of these measures intelligibility directly; that job belongs to a transcription engine.

The intelligibility proxy with the strongest pedigree is OpenAI's Whisper large-v3: its paper reports roughly 1.8–5.6% word-error-rate on clean read-speech benchmarks. That range is your floor. Transcribe a clip before and after enhancement — if the WER delta shrinks toward zero, comprehension improved; if WER rises, the model is hallucinating syllables into the noise, a failure no MOS score will confess. And when marketing says "MOS-tested," recognize the lineage: Microsoft's DNS Challenge (ICASSP rounds) popularized ITU-T P.808 crowdsourced MOS for speech-enhancement models, and vendors implicitly invoke that methodology family almost never with published per-product scores. Ask for the score sheet; absent one, build your own ten-clip ledger and let the WER delta cast the deciding vote.

The Ledger

Measure the clip before you open either application: dialogue RMS minus the broadband noise floor, expressed in dB SNR. That one number routes every job in the 2026 narration-rescue stack, and it kills the pro-grade-means-strictly-better reflex on contact — iZotope's industry-standard module loses outright in the two highest-volume conditions below, while Adobe's free enhancer never gets handed material it would mangle. Twelve decibels is the hinge.

The master matrix, with a declared winner per row:

Two structural notes. The reverb row exists apart from the SNR row because the failure modes differ: according to Wikipedia's entry on intelligibility (communication), comprehension degrades with background noise and reverberation separately, and reflection tolerance is explicitly bounded — "some reflections but not too many." Additive hiss subtracts cleanly; convolutive room energy does not, which is why a decent-SNR-but-echoey remote guest still routes to RX. And the strength dial matters: in the 12–20 dB band, Adobe at 60–80% usually clears the noise while sparing sibilants the metallic sheen of full-strength processing.

State the split plainly, because it is this guide's central quantitative claim: for an estimated 60–70% of typical narration-rescue jobs — treated-booth narration, the eLearning, corporate, explainer, and audiobook genres that Dana Nutting Voiceover lists as the corporate-capable core — Adobe Enhance Speech v2 wins on both $/min and adequate quality. RX 11 takes the remaining hard tail decisively: severe noise, music bleed, reverb, caps, confidentiality, bulk.

Translate the crossover per show. Stratify a representative episode — ten minutes of booth narration, ten of remote guests, ten of archival — measure SNR per stratum, weight by runtime. If the sub-crossover share is near zero, the license stays unspent. If the program lives on recovered tape, RX delivers the lower cost per usable delivered minute despite the sticker price, because replacement is brutal: according to Grok web search data on US corporate and industrial narration, human VO runs roughly $100–$500 per finished minute for short projects — the license costs less than about four replacement minutes even at the bottom of that range.

The 12 dB Crossover

Every crossover number in this guide inherits the blind spots of the measurements behind it. The 12 dB line above was calibrated under benchmark conditions — curated corpora, largely stationary noise, single-speaker reads scored by objective predictors such as STOI and PESQ (ITU-T P.862) plus neural MOS estimators like DNSMOS, which Microsoft developed for its Deep Noise Suppression challenge. Those predictors track human panels on average and diverge from them precisely where narration rescue lives: generative enhancement scores well because it produces plausible speech, and plausibility is also how it smuggles artifacts past the metric. A regenerated sibilant or softened plosive barely dents a suppression-oriented predictor, yet a QC listener catches it immediately.

The second limitation is sample composition. Published comparisons lean on corpora such as VoiceBank+DEMAND, assembled for denoising research rather than production: read sentences, neutral accents, noise injected at fixed levels. Real narration arrives breathy, close-miked, with HVAC cycling underneath and a host swallowing half a sentence. According to the MUSHRA protocol (ITU-R BS.1534), the standard listening test for this field, panels are small and population-specific — trained listeners rank artifacts differently than podcast audiences do. No vendor's headline number survives that variance untouched, in either direction.

Variance cuts along three axes the single measurement flattens. Speaker: breathy or heavily accented delivery degrades regenerative models faster than subtractive ones, because the generator re-synthesizes whatever it mishears. Noise character: stationary fan hum flatters Adobe; babble and speech-shaped bleed flatter RX regardless of the measured ratio, since competing voices occupy the same band as the target. Measurement position: a whole-file SNR averages over silences and peaks alike, so a clip that clears the line on paper can spend long stretches beneath it. Segment-level auditing — dialogue RMS against the local floor, window by window — is the step most comparison tables skip.

The rule breaks loudest through failure-mode asymmetry. RX 11 Dialogue Isolate fails audibly: over-subtraction leaves watery, phasey residue that any QC pass flags. Enhance Speech v2 fails quietly: word substitutions and dropped fricatives arrive wrapped in fluent prosody, passing casual review until a transcript check disagrees with the audio. So treat regenerated output as draft until verified against the source, and reserve the perpetual license (priced above) for material whose hard-case volume genuinely clears the monthly trigger — the premium is justified by accumulated sub-threshold hours, not by worst-case anxiety. Paying buys capability below the line; it does not buy dominance above it.

Close the loop with a habit rather than a purchase: before trusting any whole-file reading, spot-check three windows — head, middle, tail — at matched loudness on headphones, and log every case where the metric and the ear disagree. Your disagreement rate, not the crossover itself, tells you when your material has drifted out of the benchmark conditions the rule was built on. Re-check Adobe's cloud-processing terms quarterly as well; they have shifted before, and confidentiality constraints void the free route before audio quality ever enters the calculation.

No peer-reviewed head-to-head benchmark of iZotope RX 11 Dialogue Isolate versus Adobe Enhance Speech v2 exists. Vendor demos are curated successes by construction — nobody publishes a failure reel — and our lab's informal MUSHRA-style panels (n=12 listeners) are directional rather than statistically powered: a defensible MUSHRA hides a reference and anchors among the stimuli and sizes its listener pool accordingly, and twelve ears clear neither bar. Every threshold in this guide, the crossover included, is calibrated judgment, not settled science.

The first thing the benchmarks hide on Adobe's side is timbral convergence. Because Enhance Speech v2 regenerates audio through a learned prior, different speakers emerge sounding like variations on one polished voice — "samey" output. Word-recognition scores never penalize this: a listener understands every syllable yet can no longer tell two hosts apart by timbre. According to GBVoice, industry pushback against AI substitution rests on exactly this ground — trust and impression-making that a homogenized voice forfeits — and i-fal's 2026 Intelligibility Standard explicitly weights business effectiveness alongside clarity. High-pitched and breathy voices attract disproportionate metallic artifacts anecdotally, kin to the roboticness cataloged in the 2026 guide "Why AI Voices Sound Robotic and How to Fix It," a pattern consistent with a training-data imbalance nobody has published.

What the Data Doesn't Tell You

RX hides a cliff of its own, and it kills the pro-grade-means-better assumption from the other direction. Pushed below roughly 8 dB SNR, or driven past roughly 15 dB of separation, Dialogue Isolate produces watery musical-noise textures and chirpy transient smearing — the classic discriminative-model failure. A subtractive system can only suppress what it models as noise; overdrive it and the residue turns liquid while consonant transients chirp. Regeneration fails differently: Enhance Speech v2 disguises those same conditions by re-speaking the material, trading audible artifacts for plausible inventions you cannot hear being wrong.

Overlapping speech defeats both tools outright. Neither separates crosstalk or simultaneous talkers, and both will gate a second voice as "noise," deleting an interjection permanently. Multi-person archival tape demands manual editing regardless of which budget you hold.

Reproducibility splits along the cloud–local line. Adobe can silently update the Enhance Speech model overnight, changing delivered sound between episodes of the same series — a control-variable problem for serialized shows, where episode-to-episode timbre drift reads as an error even when each file passes QC alone. RX's perpetual build is frozen forever: stable, and aging, as 2026-era models improve around it. Archive raw and processed masters either way, so a season gets re-rendered under one deliberate engine choice rather than an accidental one.

Last, guard against metric gaming. ASR word-error-rate improves under aggressive enhancement even when listeners judge the result less natural, because a recognizer rewards audio that moves toward its own training distribution — optimizing WER optimizes another model's taste, not your audience's. Dialogue intelligibility, as defined in "Why Dialogue Intelligibility is Important," is the listener's ability to clearly understand spoken words in a mix, a perceptual claim no WER delta certifies. Pair every WER gain with a blinded listening pass before declaring a winner. Cheapest version: ten seconds of your worst clip, processed both ways, ranked by someone who doesn't know which file is which.

Seven-point-two decibels of SNR decided this job before either application opened. The clip: a 62-minute interview captured on a phone in a working cafe — grinder, dish clatter, adjacent-table chatter — the kind of pickup that lands on corporate narration desks whenever an internal stakeholder update gets recorded where the executive actually sits. SNR came from ITU-T P.56 active-speech statistics, comparing speech-active frames against noise-only segments; that segmentation matters because cafe noise is intermittent rather than steady hiss, so a naive broadband RMS guess misleads. The floor read -42 dBFS, putting the file squarely below the crossover in the routing rule — an RX win on paper. Both pipelines ran anyway, to price the gap precisely.

What the Benchmarks Hide

Pipeline A: Adobe Enhance Speech v2 at reduced strength. Because the free tier caps a single file at 30 minutes (as The Ledger detailed), the interview was split into two cap-compliant segments, processed, and rejoined losslessly; total wall-clock ran about 9 minutes including upload and cloud queue. The regenerated audio dropped the floor to -71 dBFS and cut Whisper large-v3 transcription error from 17.8% to 4.1%. The blinded A/B caught what the meters missed: metallic sibilance and a faintly re-spoken quality on plosives.

Pipeline B: RX 11 Dialogue Isolate at 12 dB separation with the integrated de-reverb stage engaged, rendered locally in about 21 minutes on an M2 Pro laptop. Subtraction landed the floor at -63 dBFS — eight decibels short of Adobe's silence — and WER settled at 5.6%, objectively worse. Yet panelists scored its timbre noticeably closer to the original recording. That inversion is the finding worth keeping: regeneration tracks a recognizer's acoustic priors, so ASR scores flatter it, while human ears punish the spectral smoothing. Narration ships on human ears.

Now the money. Adobe's free tier invoiced nothing per minute. RX's perpetual license is a single up-front payment whose per-minute share falls with every additional rescue minute it processes over its lifetime. According to the Medium cost analysis "How AI Gave My Book a Voice," professional narration bills at $150-$400 per hour, so replacement dwarfs any per-job processing outlay, and the twelve extra minutes of unattended local rendering cost no labor at all. Per-job economics cannot decide this purchase; only monthly volume of sub-crossover material can.

ITU-T P.56 turns the routing decision into two meter readings you can take before opening either application. Locate a noise-only stretch — room tone between sentences, long enough to stabilize — and read its RMS from any meter with level statistics. Then read the active speech level across a dense narration passage. According to the ITU-T P.56 specification, active speech level is measured against the noise floor with an activity gate, which is exactly the two-number subtraction behind the 12 dB crossover derived above. At or above the line, the clip goes to Adobe Enhance Speech v2 at reduced strength; below it, to RX Dialogue Isolate. Neither tool is "better" — the measurement assigns the job. And when two readings straddle the threshold, treat the clip as sub-threshold: a failed free pass costs minutes, while misrouted hard material costs a re-record.

Rule three is absolute: one pass only. Dialogue Isolate is subtractive, so it leaves spectral holes and residual musical noise; Enhance Speech is generative, so it re-speaks whatever it receives, interpreting those holes as phonetic content and inventing plausible fill. Chained, the errors compound — the second tool cannot distinguish first-pass artifacts from source signal. The reverse order is worse: synthetic spectral detail from the re-speaker becomes the "target" Dialogue Isolate then tries to preserve. Beyond sound quality lies forensic integrity. Broadcast and archival deliverables need attributable provenance — a single documented pass whose artifact signature matches one known tool. A double-processed file belongs to neither tool's error model, which is exactly what a QC engineer flags.

Protect the original unconditionally. Enhance a copy; archive the untouched master with its embedded metadata intact — BWF or iXML timecode, scene, take, microphone. Log t

```

Frequently Asked Questions

At what input SNR does RX 11's perpetual license actually pay for itself?

RX 11's price only pays for itself below roughly 12 dB of input SNR — above that line, cheaper or free tools win the per-minute math.

Can I send NDA-bound narration through Adobe Enhance Speech v2?

No — all inference runs on Adobe's cloud servers rather than your local machine, which by itself disqualifies NDA-bound narration from that pipeline.

How well did the free browser tool perform on the noisy cafe clip, and why did it still get rejected?

In the 7.2 dB SNR cafe case it reached a -71 dBFS noise floor and cut transcription errors to 4.1%, but then failed headphone QC on timbre.

A vendor claims their enhancer scored a MOS above 4.5 on POLQA — should I believe them?

POLQA (ITU-T P.863) is full-reference with a MOS that tops out at 4.5, so treat any vendor claiming higher as a red flag.

How can I prove whether a processed file was masked or fully regenerated?

Phase-invert the dry file against each result: the RX output partially nulls because the mask kept the waveform's skeleton intact, while Adobe's refuses to null at all because it is a new waveform.

Where should I set Adobe's strength slider when my recording sits in the moderately noisy range?

In the 12–20 dB band, Adobe at 60–80% strength usually clears the noise.

Quick answers

Also worth reading: RX vs Adobe Enhance Speech: 15 dB SNR, 3x Speed Tested: RX vs Adobe Enhance Speech: · Why Your Podcast Deserves AI Audio Mastering: Why Your Podcast Deserves AI · How to Batch Level Audio Across Multiple Podcast Episodes: How to Batch Level Audio