GigaAM: Dual Advances in Speech and Photonics

• GigaAM is a dual-domain innovation that integrates Conformer-based self-supervised speech recognition with gigahertz-frequency acousto-optic photonic modulation for scalable systems.
• The speech framework employs dynamic chunkwise attention and HuBERT-style pretraining to achieve up to 50% relative word error rate reduction compared to established ASR baselines.
• The photonic platform leverages optimized device architectures on a SiNₓ–AlN stack to enable efficient, high-speed phase modulation at visible wavelengths with dramatically reduced power consumption.

GigaAM denotes two distinct technological advances: a family of Conformer-based self-supervised pretraining frameworks for speech recognition (Kutsakov et al., 1 Jun 2025) and a gigahertz-frequency acousto-optic phase modulation platform for integrated photonics (Freedman et al., 11 Feb 2025). Both share the “GigaAM” designation but address fundamentally different research domains: one in large-scale speech foundation models, the other in visible-light acousto-optic modulation for scalable quantum and microwave photonic systems.

1. GigaAM for Self-Supervised Speech Recognition

Model Architecture and Pretraining

The GigaAM speech encoders are based on the Conformer backbone, where each layer incorporates multi-head self-attention, convolution modules, feed-forward sublayers, and layer normalization. In the canonical 240 M parameter version, there are $L=18$ layers, each featuring $H=8$ attention heads with hidden size $d=512$ and a convolutional kernel ($K=5$ for a 200 ms receptive field). The pretraining objective follows HuBERT-style masked-language modeling: for masked frames $M\subset\{1,...,T\}$ of encoder outputs $x=(x_1,\dots,x_T)$, GigaAM predicts targets $z_t$ derived from a fixed, supervised CTC-based ASR teacher. The loss is

$$L_{\text{MLM}} = -\sum_{t\in M} \log p(z_t \mid x_{\neg M}),$$

where teacher-derived targets $z_t$ are 1024-way cluster assignments from K-means over final-layer CTC encoder activations, providing semantically rich labels at the phoneme/word level rather than low-level audio quantizer codes (Kutsakov et al., 1 Jun 2025).

Chunkwise Attention and Context Adaptivity

GigaAM employs chunkwise attention to restrict computation during pretraining and enable both streaming and full-context inference. For queries at time $t$ in chunk $H=8$0 of length $H=8$1, attention is limited to

$H=8$2

The model samples $H=8$3 uniformly from $H=8$4 at pretraining, decoupling model adaptation from downstream streaming or long-form ASR. This dynamic chunking removes the need for separate pretraining for streaming and full-context modes.

Scaling Laws

Scaling experiments on Russian speech show that increasing the quantity of unlabeled audio beyond $H=8$5 hours provides negligible returns in word error rate (WER) reduction. Model performance saturates around 100 M parameters; at this scale, GigaAM outperforms its 240 M CTC teacher baseline (4.21% vs 4.62% WER). In smaller models (30–100 M), CTC regularization yields benefits when trained from scratch, but these gains are marginal ($H=8$6) after GigaAM pretraining.

Russian ASR Benchmarks

On Russian ASR benchmarks, GigaAM-based models achieve substantial WER reductions relative to large-scale models such as Whisper-large-v3:

Model	Golos Farfield	Russian MCV-19	Russian LibriSpeech
XLS-R (finetuned)	15.7%	8.0%	16.1%
Whisper-large-v3	16.6%	5.5%	9.5%
FastConformer-RNNT	6.6%	5.7%	11.3%
Ours (CTC)	4.3%	3.1%	5.5%
Ours (RNNT)	3.9%	2.7%	5.5%

Relative to Whisper-large-v3, GigaAM achieves up to 50% relative WER reduction (Kutsakov et al., 1 Jun 2025).

Open-Source Resources

All foundation models, ASR fine-tuned checkpoints (CTC and RNNT), and inference toolkit are released under the MIT license at https://github.com/salute-developers/gigaam.

Technical Insights

Key contributors to GigaAM’s improvements over prior SSL include semantically enriched pretraining targets, unified training for context length, and efficient Conformer modules with tunable receptive fields. Ablation studies confirm robust generalization with dynamic chunking and performance tradeoffs with convolution kernel size.

2. Gigahertz-Frequency Acousto-Optic Modulation Platform (GigaAM) in Integrated Photonics

Physical Principles of GigaAM

Gigahertz-frequency acousto-optic modulation utilizes photoelastic and moving-boundary effects within a monolithic photonic circuit. An RF-driven piezoelectric transducer excites a localized breathing-mode mechanical resonance. The resulting strain field $H=8$7 and displacement $H=8$8 modulate the permittivity tensor and induce refractive index oscillations:

$H=8$9

$d=512$0

The optical phase of a guided mode is modulated at drive frequency $d=512$1, with depth $d=512$2.

Device Architecture

The GigaAM modulator is fabricated on a 200 mm silicon wafer. The stack comprises a 1 µm AlN piezoelectric layer between Mo electrodes, a 400 nm SiN$d=512$3 waveguide with 2 µm SiO$d=512$4 cladding, and an undercut "island" structure on a nanopillar array. Co-location of mechanical and optical resonances is achieved by placing the SiN$d=512$5 waveguide atop the AlN transducer. Key geometrical parameters are optimized to maximize optomechanical overlap, with w$d=512$6=1.25 µm (oxide pad), w$d=512$7=0.5 µm (waveguide core), and 400 nm SiN$d=512$8 thickness (Freedman et al., 11 Feb 2025).

Device Performance and Parameter Relationships

On-resonance, the device presents $d=512$9 and couples $K=5$0 of microwave power from a 50 $K=5$1 line. With $K=5$2 mW input, $K=5$3 V drives the transducer, yielding modulation depth $K=5$4 rad at 2.31 GHz, limited by nonlinearities at high drive. The $K=5$5-phase voltage-length product is $K=5$6 V·cm at 730 nm. This figure of merit constitutes a 15$K=5$7 reduction relative to thin-film lithium niobate devices, while required drive power for $K=5$8 rad is reduced by $K=5$9100$M\subset\{1,...,T\}$0.

Experimental Characterization

A fiber-based interferometer with an acousto-optic frequency shifter (AOFS) provides heterodyne detection of phase modulation sidebands, enabling precise extraction of $M\subset\{1,...,T\}$1 from measured beat note powers $M\subset\{1,...,T\}$2. S-parameters (reflection $M\subset\{1,...,T\}$3) yield mechanical $M\subset\{1,...,T\}$4 and resonance frequency. Drive sweeps from 0 to 15 mW confirm alignment with Bessel-function sideband physics for phase modulation up to the nonlinear threshold. Optical insertion loss is $M\subset\{1,...,T\}$520 dB total (grating+propagation), with established SiN$M\subset\{1,...,T\}$6 power handling up to hundreds of milliwatts at 729 nm.

Comparison to Prior Art and Applications

Platform	$M\subset\{1,...,T\}$7	RF Power ($M\subset\{1,...,T\}$8 rad)	Optical Power Handling
Bulk LN (free-space)	$M\subset\{1,...,T\}$920 V	100s mW–W	$x=(x_1,\dots,x_T)$01 mW (damage)
Waveguide LN	$x=(x_1,\dots,x_T)$13–5 V·cm	30–100 mW (MgO-doped)	10s mW
GigaAM (SiN$x=(x_1,\dots,x_T)$2–AlN)	0.26 V·cm	15 mW ($x=(x_1,\dots,x_T)$3 rad), $x=(x_1,\dots,x_T)$4 V$x=(x_1,\dots,x_T)$5	$x=(x_1,\dots,x_T)$6100 mW (730 nm)

GigaAM devices satisfy requirements for quantum control in trapped-ion and neutral-atom platforms: GHz-range phase modulation, high extinction, sub-µs switching (ring-up/decay $x=(x_1,\dots,x_T)$7 ns), and robust power handling. Wafer-level integration enables thousands of independently driven channels for multi-beam quantum control and microwave photonic signal processing (Freedman et al., 11 Feb 2025).

Evolution and Future Directions

Simulations suggest straightforward extension to operation from 1–5 GHz and 400–1000 nm wavelengths by varying device cross-section and cladding. The same process will yield devices for atomic cooling (780, 422, 313 nm), and targeted GHz frequency shifts (0.3–13 GHz). Proposed next steps include single-sideband operation, nonreciprocal Brillouin on-chip isolators, and fully integrated photonic mesh platforms for qubit control.

3. Distinguishing “GigaAM”: Two Unrelated Domains

Despite nominal overlap in abbreviation, the “GigaAM” research lines in self-supervised speech pretraining (Kutsakov et al., 1 Jun 2025) and acousto-optic integrated photonics (Freedman et al., 11 Feb 2025) are independent in concept and application. Contextual disambiguation is necessary; in speech recognition literature, “GigaAM” refers explicitly to Conformer SSL models, while in photonics it uniquely identifies the CMOS-integrated gigahertz acousto-optic modulator platform.

4. Technical Contributions and Impact

The GigaAM speech framework demonstrates that self-supervised Conformer architectures, when paired with high-level, teacher-derived clustering targets and context-adaptive pretraining, deliver state-of-the-art Russian ASR performance with highly efficient scaling. Its unified chunking approach obviates separate models for streaming and full-context modes, and the open-source release under an MIT license catalyzes adoption.

The GigaAM integrated photonic modulator establishes a new reference standard for modulation efficiency at visible wavelengths, reducing both required voltage and RF power orders-of-magnitude below prior art. Its compatibility with high optical powers and wafer-scale manufacturability uniquely position it for quantum-classical hybrid systems and next-generation microwave photonic circuits.

5. Future Research Directions

Key open directions for GigaAM speech frameworks include expansion to other languages, integration with multilingual and multimodal models, and further analysis of optimal cluster target construction for unsupervised learning. For GigaAM photonic modulators, research will focus on sideband-selective modulation, complex-valued photonic network synthesis, cryogenic compatibility for quantum device integration, and in-situ control mesh architectures.

• “GigaAM: Efficient Self-Supervised Learner for Speech Recognition” (Kutsakov et al., 1 Jun 2025)
• “Gigahertz-Frequency, Acousto-Optic Phase Modulation of Visible Light in a CMOS-Fabricated Photonic Circuit” (Freedman et al., 11 Feb 2025)