qVocalTract: Efficient encoding framework of vocal-tract spectra on a quantum computer
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.
Rosa: I'm Rosa.
Dev: And I'm Dev, here with Taro.
Rosa: We're looking at the paper: "qVocalTract: Efficient encoding framework of vocal-tract spectra on a quantum computer".
Dev: Here's the gist of the paper. The qVocalTract framework proposes an efficient encoding method for vocal-tract transfer functions on quantum computers by using optimized, ancilla-free staircase layers to preserve acoustic structure under hardware noise The comparison between this approach and sequential matrix-product-state preparation reveals a tradeoff between approximation accuracy and accumulated circuit noise Encoding and State Preparation The paper details the encoding process where the vocal-tract transfer function is stored by sampling it at a specific frequency, resulting in an encoded state ψ⟩ The target amplitudes are nonpositive, and the state is represented using a right-canonical tensor-train singular-value decomposition truncated at bond dimension χ Sequential preparation involves extending each site isometry to a unitary on one data qubit and ⌈log2 χ⌉ ancilla qubits Ancilla-Free Staircase Optimization The core contribution is the ancilla-free staircase preparation, which constructs the layers U1 through UL by refining gates with local polardecomposition updates This optimization involves writing the target overlap for one gate g as Tr(gE) and finding the maximizing unitary g = V W† The fidelity is measured by ⟨ψϕ⟩2 Performance Comparison Across Configurations The comparison across different preparation methods shows that the three-layer staircase provides the best hardware accuracy for the reference vowel and improves mean accuracy across 44 configurations Specifically, raw-count recalculation gives mean RMSEs of 4.77, 7.16, and 9.22 dB for three layers and sequential χ = 2, 4 The staircase again wins all 44 comparisons against each baseline Modal Probe Coherence A separate modal-probe experiment tests coherent processing by using square-root-magnitude encoding rather than the log-magnitude state in Eq.
Rosa: (1) Correlation with the ideal prepared state probe is 0.974 This experiment tests coherent processing of the prepared state, rather than speech synthesis The results connect state-preparation cost to acoustic accuracy, showing that staircase layers offer a useful balance between approximation accuracy and accumulated noise on the tested processor Extending this approach to broader physical vocal-tract models and coherent acoustic operations is the next step toward quantum speech processing Table II TABLE II ALL 44 CASES ON I B M K A W A S A K I, WITH RECORDED READOUT CORRECTION: MEAN 256-BIN RMSE (DB), MAE OF DETECTED RESONANCES (HZ), AND THE FRACTION WITH ALL THREE WITHIN 100 HZ (“OK”).
Dev: “2Q” IS THE MEAN COMPILED GATE COUNT ROUNDED TO THE NEAREST INTEGER; 20 000 SHOTS PER CIRCUIT.
Rosa: Table III TABLE III REFERENCE-VOWEL 256-BIN RMSE (DB) ON THE IDEAL SAMPLER, STATIC FA K ETO R I N O NOISE MODEL, AND HARDWARE.
Title and authors: Rosa: So, looking at this paper's title, "qVocalTract: Efficient encoding framework of vocal-tract spectra on a quantum computer," it tells us right away that they are tackling the problem of encoding these complex vocal tract sounds onto a quantum system.
Dev: It’s not just about making it fit; it’s about finding an efficient framework that handles the physical reality of how speech sounds are shaped by the vocal tract, and they propose using optimized, ancilla-free staircase layers for spectral loading.
Taro: That idea of using staircase layers to handle the noise is interesting because typically you need extra qubits or more complex circuits just to manage that kind of noise.
Rosa: Right, and what they show with their simulations on a superconducting processor is that this staircase preparation lowers the spectral error and improves resonance recovery compared to sequential matrix-product-state preparation.
Dev: The comparison across different configurations shows that the three-layer staircase provides the best hardware accuracy for the reference vowel and improves mean accuracy across forty four configurations, specifically raw count recalculation gives mean root mean square errors of four point seven seven, seven point one six, and nine point two two decibels for three layers and sequential chi equals two or four.
Taro: So they're showing that this staircase method wins all forty four comparisons against each baseline they tested.
Rosa: And then there’s this separate modal probe experiment where they test coherent processing by using square-root-magnitude encoding instead of the log-magnitude state in equation one, and it actually shows a correlation with the ideal prepared state probe of zero point nine seven four.
Dev: So that suggests that once you prepare the state this way, you can actually use it for coherent operations rather than just focusing on speech synthesis.
The paper's summary: Rosa: To summarize what they are doing with the qVocalTract framework, they detail the encoding process where the vocal-tract transfer function is stored by sampling it at a specific frequency, resulting in an encoded state psi.
Dev: The target amplitudes for this state are nonpositive and they represent the state using a right-canonical tensor-train singular value decomposition truncated at bond dimension chi.
Taro: So basically, you take the acoustic data from speech and turn it into a quantum representation that’s structured in this specific tensor format, which is then prepared on the quantum computer.
Rosa: The sequential preparation involves extending each site isometry to a unitary on one data qubit and ceiling log base two chi ancilla qubits.
Dev: Their core contribution is the ancilla-free staircase preparation, which constructs the layers u sub one through u sub L by refining gates with local polardecomposition updates.
Taro: So it's an optimization technique to build up those layers without needing those extra ancilla qubits, which saves resources on the hardware side.
Rosa: This optimization involves writing the target overlap for one gate g as the trace of g times e and then finding the maximizing unitary g equals v w dagger.
Dev: The fidelity is measured by the absolute value of inner product psi phi squared over two.
The paper's improvements: Rosa: What they found in terms of improvements is that this ancilla-free staircase preparation improves spectral recovery over the tested sequential baselines, which leads to better preservation of acoustic structure under hardware noise.
Dev: Specifically, the three layers provide the best hardware accuracy for the reference vowel and improve mean accuracy across forty four configurations.
Taro: They also found that raw count recalculation gives mean root mean square errors of four point seven seven, seven point one six, and nine point two two decibels for three layers and sequential chi equals two or four.
Rosa: The staircase again wins all forty four comparisons against each baseline they tested, which is pretty significant because it shows a clear balance between approximation accuracy and accumulated circuit noise on the tested processor.
Dev: The results connect state preparation cost directly to acoustic accuracy, showing that staircase layers offer a useful balance between approximation accuracy and accumulated noise on the tested processor.
Conclusion: Rosa: So to wrap up, this paper shows that the ancilla-free staircase preparation is a way to improve spectral recovery over sequential baselines.
Dev: They found that three layers provide the best hardware accuracy for the reference vowel and improve mean accuracy across forty four configurations.
Taro: It’s showing us how to choose preparation circuits based on acoustic error and resonance recovery, which gives us some practical criteria for choosing these circuits in quantum acoustic processing.
Rosa: The overall implication is that this method offers a useful balance between approximation accuracy and accumulated noise on the tested processor, guiding practical support for further quantum acoustic processing with physical vocal tract models.
Dev: This finding connects state preparation cost directly to acoustic accuracy, which is important because it shows how to manage those trade offs when implementing these kinds of quantum encoding schemes based on physical vocal tract models.
Taro: We can use this to guide the depth of sequential preparation layers by comparing noise models and hardware results, which informs practical choices for implementing quantum acoustic encoding schemes on physical vocal tract models.
Rosa: And we’re looking forward to extending this approach to broader physical vocal tract models and coherent acoustic operations as the next step toward quantum speech processing.
Sheng Li, Jiyi Li
Graduate School of Informatics, Kyoto University · School of Engineering, Institute of Science Tokyo · TRIP, RIKEN
eess.AS
Submitted: 2026-11-11
Updated: 2026-11-11
Comments: under review
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 92/100
The gist: The qVocalTract framework proposes an efficient encoding method for vocal-tract transfer functions on quantum computers by using optimized…
Key concepts
- Vocal-Tract Transfer Function
- This is a mathematical representation of how sound waves change as they pass through the vocal tract (the tube of air in the body). The paper encodes this function into a quantum state to allow for quantum computation related to speech analysis.
- Ancilla-Free Staircase Optimization
- This is the core technique used to build the quantum circuit layers without needing extra helper qubits (ancillas). By refining gate operations based on local updates, it constructs the required unitary transformations efficiently while minimizing noise accumulation during preparation.
- Matrix Product State (MPS) Decomposition
- This is a way to represent large, complex quantum states—like those encoding vocal tract data—using a compact structure called a tensor-train decomposition. This allows the state to be stored and manipulated using fewer resources on quantum hardware.
- Modal Probe Coherence
- This experiment tests how well the prepared quantum state can handle coherent processing, rather than just synthesizing speech. Using square-root-magnitude encoding, it shows high correlation with an ideal probe, indicating the state is suitable for testing complex quantum operations.
Terminology
Summary
The qVocalTract framework proposes an efficient encoding method for vocal-tract transfer functions on quantum computers by using optimized, ancilla-free staircase layers to preserve acoustic structure under hardware noise The comparison between this approach and sequential matrix-product-state preparation reveals a tradeoff between approximation accuracy and accumulated circuit noise
Encoding and State Preparation
The paper details the encoding process where the vocal-tract transfer function is stored by sampling it at a specific frequency, resulting in an encoded state ψ⟩ The target amplitudes are nonpositive, and the state is represented using a right-canonical tensor-train singular-value decomposition truncated at bond dimension χ Sequential preparation involves extending each site isometry to a unitary on one data qubit and ⌈log2 χ⌉ ancilla qubits
Ancilla-Free Staircase Optimization
The core contribution is the ancilla-free staircase preparation, which constructs the layers U1 through UL by refining gates with local polardecomposition updates This optimization involves writing the target overlap for one gate g as Tr(gE) and finding the maximizing unitary g = V W† The fidelity is measured by ⟨ψϕ⟩2
Performance Comparison Across Configurations
The comparison across different preparation methods shows that the three-layer staircase provides the best hardware accuracy for the reference vowel and improves mean accuracy across 44 configurations Specifically, raw-count recalculation gives mean RMSEs of 4.77, 7.16, and 9.22 dB for three layers and sequential χ = 2, 4 The staircase again wins all 44 comparisons against each baseline
Modal Probe Coherence
A separate modal-probe experiment tests coherent processing by using square-root-magnitude encoding rather than the log-magnitude state in Eq. (1) Correlation with the ideal prepared state probe is 0. p. 4 )
Square-root-magnitude encoding | correlation = 0. 974
Target-state probe
Prepared-state probe
Hardware
Fig.Where the paper says it
Conclusion and Future Directions
The results connect state-preparation cost to acoustic accuracy, showing that staircase layers offer a useful balance between approximation accuracy and accumulated noise on the tested processor Extending this approach to broader physical vocal-tract models and coherent acoustic operations is the next step toward quantum speech processing
Table II
TABLE II
ALL 44 CASES ON I B M K A W A S A K I, WITH RECORDED READOUT CORRECTION: MEAN 256-BIN RMSE (DB), MAE OF DETECTED RESONANCES (HZ), AND THE FRACTION WITH ALL THREE WITHIN 100 HZ (“OK”). “2Q” IS THE MEAN COMPILED GATE COUNT ROUNDED TO THE NEAREST INTEGER; 20 000 SHOTS PER CIRCUIT. p. 3 4
TABLE II
ALL 44 CASES ON I B M_K A W A S A K I, WITH RECORDED READOUT CORRECTION: MEAN 256-BIN RMSE (DB), MAE OF DETECTED RESONANCES (HZ), AND
THE FRACTION WITH ALL THREE WITHIN 100 HZ (“OK”). “2Q” IS THE MEAN COMPILED GATE COUNT ROUNDED TO THE NEAREST INTEGER; 20 000 SHOTS
PER CIRCUIT.Where the paper says it
Table III
TABLE III
REFERENCE-VOWEL 256-BIN RMSE (DB) ON THE IDEAL SAMPLER, STATIC FA K ETO R I N O NOISE MODEL, AND HARDWARE. PARENTHESES GIVE THE GATE COUNT COMPILED FOR THE NOISE MODEL. p. 3 TABLE III
REFERENCE-VOWEL 256-BIN RMSE (DB) ON THE IDEAL SAMPLER, STATIC
FA K ETO R I N O NOISE MODEL, AND HARDWARE. PARENTHESES GIVE THE
GATE COUNT COMPILED FOR THE NOISE MODEL.Where the paper says it
Table I
TABLE I
REFERENCE VOWEL ON I B M K A W A S A K I: 256-BIN ENVELOPE RMSE (DB), INCLUDING THE APPROXIMATION COLUMN, AND DETECTED RESONANCES (HZ). TARGET RESONANCES ARE 527, 1113, AND 2285 HZ. p. 1 The reference case is s1-01, tense /a/ from speaker 1. Its target resonances are 527, 1113, and 2285 Hz. TARGET RESONANCES ARE 527, 1113, AND 2285 HZ. “2Q” DENOTES THE COMPILED CZ COUNT. “2Q” DENOTES THE COMPILED CZ COUNT. RECOVERED VALUES USE THE RECORDED
READOUT CORRECTION.Where the paper says it
Where the paper says it
Where the paper says it
Figure 3
Fig. 3. 44-case sweep with the recorded readout correction
Figure 4
Fig. 4. Modal probe with square-root-magnitude encoding and a one-layer staircase
Section VI Conclusion
In this paper, the ancilla-free staircase preparation improves spectral recovery over the tested sequential baselines Three layers provide the best hardware accuracy for the reference vowel and improve mean accuracy across 44 configurations Acoustic error and resonance recovery provide practical criteria for choosing preparation circuits
Section V Discussion
The results connect state-preparation cost to acoustic accuracy, showing that staircase layers offer a useful balance between approximation accuracy and accumulated noise on the tested processor Classical optimization uses the full target vector for each case, and envelope recovery requires many samples Peak extraction is sensitive to missing and spurious resonances, so repeated hardware runs are needed to estimate variability
Section D Separate modal-probe experiment
The modal-probe experiment uses square-root-magnitude encoding rather than the log-magnitude state in Eq. (1) Linear frequency spacing implements diagonal evolution with single-qubit phase gates Controlled phase gates add entangling operations after preparation With one staircase layer, the hardware experiment uses up to 84 two-qubit gates per circuit, 48 time samples at 48 kHz, and 2000 shots per sample Correlation with the ideal prepared state probe is 0. p. 3 (3)
Controlled phase gates add entangling operations after preparation. With one staircase layer, the hardware experiment uses up
to 84 two-qubit gates per circuit, 48 time samples at 48 kHz,
and 2000 shots per sample. (3)
Controlled phase gates add entangling operations after preparation. With one staircase layer, the hardware experiment uses up
to 84 two-qubit gates per circuit, 48 time samples at 48 kHz,
and 2000 shots per sample.Where the paper says it
Where the paper says it
Index Terms
Index Terms—Quantum computing, quantum state preparation, vocal-tract transfer functions, matrix product states, acoustic modeling
References
[3] S. Li and J. Li, “Matrix-product-state compression of quantum-encoded vocal-tract transfer functions,” JASA Express Letters, vol. p. 4 1000–1010, 2006. Li, “Matrix-product-state compression of quantum-encoded
vocal-tract transfer functions,” JASA Express Letters, vol.Where the paper says it
[4] P. Birkholz, S. p. 4 094803, 2026. Birkholz, S. Birkholz, S. Kürbis, S. Kürbis, S. Häsner, R. Häsner, R. Blandin, and M. Blandin, and M. Fleischer,
“Printable 3D vocal tract shapes from MRI data and their acoustic and
aerodynamic properties,” Scientific Data, vol.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[5] G. Vidal, “Efficient classical simulation of slightly entangled quantum computations,” Physical Review Letters, vol. p. 4 Fleischer,
“Printable 3D vocal tract shapes from MRI data and their acoustic and
aerodynamic properties,” Scientific Data, vol. Vidal, “Efficient classical simulation of slightly entangled quantum
computations,” Physical Review Letters, vol.Where the paper says it
[6] I. V. Oseledets, “Tensor-train decomposition,” SIAM Journal on Scientific Computing, vol. p. 4 147902, 2003. Oseledets, “Tensor-train decomposition,” SIAM Journal on Scientific
Computing, vol.Where the paper says it
[7] C. Schön, E. Solano, F. Verstraete, J. p. 4 2295–2317, 2011. Verstraete, J. Verstraete, J. Cirac, and M. Cirac, and M. Wolf,
“Sequential generation of entangled multiqubit states,” Physical Review
Letters, vol.Where the paper says it
Where the paper says it
Where the paper says it
[8] S.-J. Ran, “Encoding of matrix product states into quantum circuits of one- and two-qubit gates,” Physical Review A, vol. p. 4 110503, 2005. Ran, “Encoding of matrix product states into quantum circuits of
one- and two-qubit gates,” Physical Review A, vol.Where the paper says it
[9] M. S. Rudolph, J. p. 4 032310, 2020. Rudolph, J. Rudolph, J. Acharya, and A. Acharya, and A. Perdomo-Ortiz,
“Decomposition of matrix product states into shallow quantum circuits,”
Quantum Science and Technology, vol.Where the paper says it
Where the paper says it
Where the paper says it
[10] M. Ben-Dov, D. p. 4 015012, 2024. Ben-Dov, D. Ben-Dov, D. Shnaiderov, A. Shnaiderov, A. Makmal, and E. Makmal, and E. Dalla Torre,
“Approximate encoding of quantum states using shallow circuits,” npj
Quantum Information, vol.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[13] A. Javadi-Abhari, M. p. 4 040326, 2021. Javadi-Abhari, M. Javadi-Abhari, M. Treinish, K. Treinish, K. Krsulich, C. Krsulich, C. Lishman,
J. Lishman,
J. Martiel, P.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[14] N Ezzell, B Pokharel, L Tewala, G Quiroz, and D A Lidar “Dynamical decoupling for superconducting qubits: a performance survey”, Physical Review Applied, vol. 20, p. 064027
[15] M Plesch and C Brukner, “Quantum-state preparation with universalˇ gate decompositions,” Physical Review A, vol. 83, p. 032302, 2011
[44] DVTD contains 44 vocal-tract configurations from two speakers, including vowels and consonants [4]
The modal-probe experiment uses square-root-magnitude encoding rather than the log-magnitude state in Eq. (1) Linear frequency spacing implements diagonal evolution with single-qubit phase gates Controlled phase gates add entangling operations after preparation With one staircase layer, the hardware experiment uses up to 84 two-qubit gates per circuit, 48 time samples at 48 kHz, and 2000 shots per sample Correlation with the ideal prepared state probe is 0.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improved state preparation circuits can be designed using ancilla-free staircase decompositions to lower spectral error and improve resonance recovery compared to sequential matrix-product-state preparation. This allows for better preservation of acoustic structure under hardware noise, leading to more accurate quantum acoustic modeling.
-
The qVocalTract framework enables the comparison of circuit cost directly with spectral envelope error and resonance recovery, providing a practical criterion for choosing state preparation circuits in quantum acoustic processing. This capability helps researchers balance approximation accuracy against accumulated circuit noise on physical processors like ibm kawasaki.
-
AI systems can perform coherent processing of vocal-tract states using modal-probe experiments with square-root-magnitude encoding, achieving high correlations (up to 0.974) with the ideal prepared state probe on hardware. This demonstrates that the prepared state is suitable for testing coherent operations rather than just speech synthesis.
-
The framework can be used to optimize the depth of sequential preparation layers by comparing noise models and hardware results, guiding which preparation depth best balances approximation accuracy and circuit cost. This informs practical choices for implementing quantum acoustic encoding schemes based on physical vocal-tract models.
Abstract
The vocal-tract transfer function links the physical articulatory model to speech. Quantum encoding of vocal-tract transfer functions requires circuits that preserve acoustic structure under hardware noise. This paper proposes the qVocalTract framework, which uses optimized, ancilla-free staircase layers for spectral loading. Simulations and a superconducting processor compare this approach with sequential matrix-product-state preparation on vocal-tract data. Staircase preparation lowers spectral error and improves resonance recovery, while a modal probe demonstrates coherent processing. The comparison reveals a trade-off between approximation accuracy and accumulated circuit noise. These findings guide practical support for further quantum acoustic processing with physical vocal-tract models.
Sources
Related papers
- WavCube: Unifying Speech Representation for Understanding and Generation via Semantic-Acoustic Joint Modeling
- From perception to production: how acoustic invariance facilitates articulatory learning in a self-supervised vocal imitation model
- Read What You Hear: Reference-Free Hypotheses Evaluation with Acoustic Discrepancy
- pyFDN: A Python Toolbox for Feedback Delay Network Design, Analysis, and Optimization
- GigaSpeechBench: A Real-World Multilingual Speech-to-Text Benchmark
- DM-ASR: Diarization-aware Multi-speaker ASR with Unified Offline and Online Decoding with Large Language Models