Cloned Voice Stereo Effects: Favor Dry Mono Over 3-Band Autopanning

Three frequency bands, each with independent motion controls: that is the PanShaper capability documented in Pluginerds’ guide. It offers an enticing way to keep lower frequencies centered while moving higher frequencies across headphones. But for cloned speech, width is not evidence of fidelity. The neural-speech question is whether the processing preserves intelligibility, speaker identity, and natural delivery when the stereo presentation disappears.

The mechanism matters more than the spectacle. ICON Collective describes Ableton Auto Pan as moving sound through opposing changes in left- and right-channel volume, not by inverting the audio waveform’s polarity. Nevertheless, mono level depends on the gain trajectories and the fold-down convention. Without those assumptions and a defined reference, a precise aggregate-loss claim is not established by the supplied sources. Decibel equations convert a specified ratio; they do not supply the missing signal model.

Favor dry mono as the baseline, then treat multiband motion as an optional effect that must justify itself. Compare processed fold-downs against the unprocessed voice, checking level separately from changes in articulation and timbre. Independent band movement can be creatively useful, but impressive headphone width should never substitute for evidence that the cloned voice remains intact in mono.

Define the Mono Reference Before Trusting a 3-Band

Before you trust a three-band autopanner on a cloned voice, build the mono reference it will be judged against. The signal path is not a synthesizer: one consented cloned-voice waveform is split by a crossover into low, middle, and high bands, each band receives independently time-varying left and right gains, and the bands are summed back to stereo. No new phonemes are generated and no neural vocoder is re-run — the sample content at the output is the same content that entered. That distinction matters because a neural synthesizer can regenerate a voice that survives mono, while a gain-motion processor can only redistribute what is already there. According to Pluginerds, PanShaper 4 exposes separate panning control per band and combines independent LFOs with envelope followers; according to ICON Collective, Ableton Live's Auto Pan uses two separate LFOs modulating left and right channel volume independently. Neither is a voice generator.

Fix the fold-down convention first: M = (L + R) / 2. Loss = 20 log10(RMS of reference mono / RMS of processed mono). Positive values mean the processed mix is quieter in mono — attenuation. A negative value means the processed mono is louder than the reference, which is a level increase, not a pass. Swapping numerator and denominator flips the sign and can turn a 1 dB failure into an apparent 1 dB success, so declare the convention before you measure anything.

The reference must be matched, not arbitrary. It needs the same crossover network, the same latency, and the same declared pan law as the moving version, with modulation disabled. If the centered reference is built with a different filter or a different center-gain convention, you are measuring the filter rather than the autopanning — and an arbitrary center gain can masquerade as mono loss in either direction.

A fourth-order Linkwitz–Riley crossover is a concrete implementation example: nominal slope of 24 dB per octave, with each branch −6 dB at its crossover frequency. Those are per-branch properties, and they do not guarantee that the complete three-way recombination is transparent. Verify the summed three-way output against the input rather than assuming separate two-way behavior composes into a clean whole.

Independent gain motion reshapes the spectral envelope over time. When the low band is attenuated while the high band is boosted, harmonic-dominant and consonant-dominant regions of the same syllable are weighted differently — timbre shifts even though no samples were removed. This is amplitude summation, not phase cancellation. The myth that mono loss must originate in phase cancellation misses the dominant mechanism here: decorrelated band gains sum to a different mono level by construction. According to Pluginerds, the supplied PanShaper 4 material documents three-band processing but reports no measured mono-fold-down loss in decibels, so treat any claimed figure as unverified until you measure it on your own material.

Measure per band, not just overall. A mix can pass the overall 1 dB limit while one band exceeds it, and that single band is where the cloned speaker's identity usually degrades first.

What BS.1770 and Speech Metrics Can

BS.1770 loudness compliance is not a mono-compatibility certificate. For cloned speech, the useful distinction is between a measurement convention and evidence that processing preserves speech. Standardized loudness and speech scores describe particular properties; none independently establishes that a moving multiband voice survives mono reproduction without damaging intelligibility or speaker identity. The centered voice remains the default unless the actual fold-down and controlled listening comparison both pass the article’s acceptance rule.

According to ITU-R BS.1770-5, the loudness algorithm assigns the left and right channels weighting factors of 1.0 when combining their frequency-weighted mean-square energies. This is energy accounting across channels, not measurement of their summed waveform. An actual mono fold-down combines waveforms before its energy is measured, introducing a cross-channel interaction term that depends on their relationship. Consequently, the stereo energy total cannot establish the folded signal’s level. Nor does every mono loss imply phase cancellation: changing channel gains can change the summed amplitude even when the channels remain phase-aligned. Independent band movement therefore guarantees neither mono stability nor greater naturalness.

According to EBU Tech, momentary loudness uses a short window and short-term loudness uses a longer window. These standardized summaries are useful for following loudness behavior, but their temporal averaging can conceal attenuation concentrated in a consonant or phoneme transition. Stronger adjacent speech can dominate the summary while a brief, linguistically important cue weakens. A momentary or short-term loudness difference should therefore be labeled by its window, not presented as the worst-case mono loss. These summaries also cannot establish that every individual band stayed within the acceptance limit.

According to IEC, the Speech Transmission Index ranges from 0 to 1. STI characterizes speech transmission through a channel using the preservation of modulation relevant to intelligibility. That purpose differs from determining whether a synthetic utterance still resembles its intended speaker. Speech can remain understandable while vocal characteristics change. For time-varying multiband processing, applicability and interpretation require particular care; an STI result must not be relabeled as a validated speaker-identity score or used to replace the controlled speech comparison.

According to ITU-T P.800, the listening-quality opinion scale runs from 1 to 5. Separate the proposed questions: “How would you rate the speech quality of this sample?” and, with an authorized target-speaker reference, “How closely does this sample resemble the reference speaker’s voice?” The latter is a separately defined resemblance judgment, not a P.800-validated identity measure. Include a distinct intelligibility check rather than assuming pleasant sound means every word remained recoverable. Use matched material and controlled presentation so a single overall preference cannot conceal reduced resemblance or word recognition.

These figures are published measurement conventions, not experimental evidence that this cloned-voice effect improves speech. Before accepting a benefit claim, require the named study, tested voices, listening conditions, and effect settings. In the evaluation record, keep fold-down measurements, intelligibility outcomes, and identity judgments separate; an attractive standardized score cannot compensate for failure in another required outcome.

Centered Dry Voice Wins Until a Moving Alternative

Centered dry voice wins because width is optional, but predictable speech preservation is not. For a cloned-voice deliverable that must survive both stereo and mono playback, movement carries the burden of proof. An untested configuration is unverified, not safe—even when its routing looks mathematically reassuring. The useful comparison is not which version sounds widest, but whether the moving version preserves intelligibility and speaker identity without exceeding the mono-loss limit established above.

Constant-sum panning changes the engineering trade rather than eliminating it. Preserving each band's left-plus-right gain can make that band's mono contribution independent of pan position, provided the sum matches the reference gain and no intervening processing breaks that relationship. Stereo energy, however, depends on how gain is distributed between channels, not merely on their sum. Independent band movement can therefore change the stereo spectral-energy balance while leaving the mono sum stable. Mathematical mono preservation does not guarantee stable stereo loudness or a preferable cloned-voice timbre.

According to Pluginerds' guide, Cableguys PanShaper 4 splits audio into three frequency bands with adjustable crossovers. That makes crossover placement part of the configuration being evaluated, not a cosmetic setting: changing it changes which speech components move together. The product architecture establishes a capability, not a speech-quality result. Independently moving bands do not necessarily make a clone more natural; nor does every mono reduction indicate phase cancellation. A position-dependent reduction in the left-plus-right gain can lower the fold-down without destructive interference.

A dry anchor introduces a parallel-path problem. Check the moving path's delay relative to the dry path, including crossover and effect latency; an apparently aligned onset does not establish matching phase response across the spectrum. Misalignment can color the combined voice rather than simply reinforce it. Even with adequate alignment, position-dependent band gains can change the wet contribution relative to the anchor, producing a shifting spectral balance. Evaluate the actual blend throughout its movement, not just at a favorable pan position. Adding dry speech can reduce modulation without removing its cause.

Keep two separate records for each candidate: the mono-preservation property its routing predicts, and the speech-quality result its rendered output demonstrates. Neither substitutes for the other. Before promoting a candidate, render the intended crossover settings, motion, and blend, then compare against the matched centered version under controlled listening conditions in stereo and mono. Require both the established overall and per-band mono-loss check and no detected loss of intelligibility or speaker identity. Until both pass, the centered dry clone remains the explicit winner.

What the Data Doesn't Tell You

A technically predictable panning effect is not an established perceptual improvement. The supplied brief contains no controlled dataset for three-band autopanning of cloned speech. Standards support measurement, and gain equations support predictions under specified signal-path assumptions; neither establishes a population-wide benefit to intelligibility, naturalness, or speaker identity. According to Pluginerds, PanShaper 4 can keep bass centered while allowing higher frequencies to move more widely. That describes a processing capability, not evidence that synthetic voices benefit from it. Independently moving bands do not necessarily make a clone sound more natural.

Attenuation and destructive interference also require different explanations. When an otherwise identical band signal reaches both channels through nonnegative amplitude gains, those gains do not themselves reverse polarity or make the copies cancel. Its mono level can nevertheless fall relative to the matched centered reference because the combined gain changes with pan position. Additional delays, all-pass processing, or decorrelation change that analysis: channel signals can then have frequency-dependent phase differences, allowing destructive interference on summation. A loss predicted from amplitude gains and a loss caused by altered channel relationships are not interchangeable diagnoses. The latter can produce spectral damage that a gain-only prediction misses.

Even a correctly measured aggregate loss is conditional on the voice and utterance. A breathy clone can place substantial energy in noise-like components; a strongly harmonic clone concentrates energy around its harmonic structure; a sibilant-heavy phrase can emphasize the upper crossover band during consonants. Identical crossover settings therefore do not imply identical exposure to the moving gains. Timing matters too: a consonant occurring near a band's attenuation maximum faces a different condition from the same consonant elsewhere in the motion cycle. A result dominated by sustained vowels cannot stand in for consonant-rich material, and an overall pass cannot excuse a failing band.

Speaker-embedding similarity leaves another evidentiary gap. An automatic encoder is not a direct readout of perceived identity: it may tolerate audible coloration while retaining a similar representation, or respond to recording-channel changes that listeners do not interpret as a different speaker. Consequently, a stable similarity score cannot establish that the clone still sounds like the same person. Conversely, a changed score does not by itself identify an identity failure. The useful distinction is between representation stability and perceptual stability; only the latter answers the identity part of the acceptance rule.

Apparent stereo preference is equally conditional. Headphones expose channel separation directly, loudspeakers introduce acoustic mixing and room effects, and a mono phone speaker removes the intended lateral presentation. A preference on one does not transfer automatically to the others. A louder rendering can also win without preserving speech better. For the controlled comparison, keep preference separate from intelligibility and identity judgments, control presentation level, and retain results by speaker, utterance, and playback condition rather than collapsing them into a single winner. These limitations narrow what a passing result establishes; they do not justify relaxing either acceptance condition or abandoning the centered default.

Worked Gain Calculation

Unit-energy panning can preserve stereo energy while losing mono level without any phase cancellation. According to Ville Pulkki’s JAES paper, “Virtual Sound Source Positioning Using Vector Base Amplitude Panning,” panning gains are normalized so their squared values sum to unity. The example below applies that published normalization principle; it is a calculation, not a reported cloned-speech experiment. Its useful distinction is between preserving the energy distributed across loudspeakers and preserving the energy of their arithmetic-average sum.

Declare three mutually uncorrelated bands, each with unit mean-square energy. At one instant, assign left/right gain pairs of (1, 0), (1/√2, 1/√2), and (0, 1) to the low, middle, and high bands, respectively. Evaluate these as frozen gain settings, with no added gain compensation, delay, or polarity inversion. Equal band energies and zero cross-band covariance are explicit assumptions—not measured speech statistics. Every pair satisfies gL² + gR² = 1, so each band retains its original summed left-plus-right energy.

For arithmetic-average fold-down, m = (L + R)/2. A band signal x therefore enters mono as [(gL + gR)/2]x. Its mono mean-square contribution is E[(gL + gR)/2]², where E is that band’s input mean-square energy. This is the calculation to inspect: the stereo normalization constrains the sum of squared gains, whereas mono energy depends on the square of their sum. Those expressions are not interchangeable.

In the matched centered reference, every band uses (1/√2, 1/√2). Arithmetic averaging therefore gives a mono amplitude coefficient of 1/√2. Squaring that coefficient makes each unit-energy band contribute 1/2. Because the bands are mutually uncorrelated, their mean-square contributions add without cross terms: the total centered-reference mono energy is 1/2 + 1/2 + 1/2 = 3/2.

For the processed setting, the low and high bands each have a mono amplitude coefficient of 1/2; the middle retains 1/√2. Their mono energy contributions are consequently 1/4, 1/2, and 1/4, totaling 1. The processed-to-reference energy ratio is 1/(3/2) = 2/3. According to Wikipedia’s “Decibel” entry, a power-ratio change uses ten times the base-ten logarithm. Applying that definition gives 10 log10(2/3) = −1.761 dB, rounded. This loss arises from gain geometry, not destructive interference.

The bandwise verdict is stricter still. Each outer band has a processed-to-reference ratio of (1/4)/(1/2) = 1/2, yielding 10 log10(1/2) = −3.010 dB, rounded; the middle band is unchanged. Reject this illustrative setting: both the overall result and the outer-band results exceed the permitted mono-loss limit. Centering wins this calculation; a favorable listening impression cannot override a failed numerical gate.

For actual speech, replace the assumed unit energies with measured band energies and retain covariance terms introduced by the crossover outputs. Do not transfer this snapshot’s result directly to an entire moving passage. A real candidate must pass the mono-loss gate and a controlled comparison detecting no loss of intelligibility or speaker identity before replacing the centered default.

How to Choose Well

The decision is not whether the effect sounds wider in stereo. It is whether the delivered stem survives the downmix the listener will actually use. Five gates, applied in order, resolve that — and the first one is the one most engineers skip.

Rule 1 — No renderer, no effect. If you cannot test through the final renderer, the actual downmix convention, or the processed stem itself, skip the effect. A stereo-only preview cannot establish compatibility of the delivered cloned voice, because the mono sum is a property of the matrix, not of the preview. The bsmith96/Reaper-Scripts documentation shows how straightforward it is to wire parameter modulation into an autopanning effect; that ease is exactly the trap. You can build a convincing stereo demo in minutes and still have no evidence about the mono fold-down. Note also that Band Barracks frames autopanning as movement and contrast for drums, percussion, and synthesizers — musical material that can absorb motion. A single cloned voice has no second element to hide behind.

Rule 2 — Reject above 1 dB. Reject any setting whose measured mono attenuation exceeds 1 dB overall or within any crossover band. Evaluate active speech across the full modulation cycle, exclude silent windows, and never normalize the processed mono file to conceal the loss — that rescales the exact quantity you are measuring. A three-band split needs three separate checks: an acceptable overall figure can mask one band that collapses while another compensates.

Rule 3 — Blinded comparison overrides level numbers. If a blinded, playback-level-matched comparison produces a repeatable intelligibility or speaker-resemblance decrement, skip the effect even when its level measurements pass. Keep the unnormalized measurements for the separate attenuation decision; the two tests answer different questions. Level matching removes loudness as a cue, which is the point — speaker identity rides on spectral envelope cues that band-dependent movement can smear.

Rule 4 — Boosts and pumping are a different failure. If the processor introduces material mono boosts or audible pumping rather than attenuation, classify that as a separate failure instead of calling a negative loss value a pass. A negative number means the mono sum got louder, which is its own defect. Reduce the processing depth or return to the centered clone.

Rule 5 — Benefit must be demonstrated, not assumed. Use the effect only when a passing configuration also demonstrates the intended spatial benefit on the target playback systems. If results are inconclusive, or the movement adds no discernible value, choose the centered version. PanShaper 4 stores up to nine custom waves per preset according to Pluginerds — more shapes mean more configurations to test, not more justification to ship one.

Apply the gates in order and stop at the first failure. The centered clone is the default deliverable; movement is the exception you earn.

What to do next