We present GigaEdge Spench, an open-source SDK for streaming speech enhancement on CPU servers, mobile devices, and in browsers. Applications can build enhancement channels from neural denoising, normalization and dynamics blocks, sample-rate conversion, and acoustic echo cancellation where supported. These stages can be combined, while normalization can also run without a neural model. The release includes an 8 kHz INT8 denoiser and discriminative and hybrid 48 kHz FP32 models. A shared C++ core supports ONNX Runtime and TensorFlow Lite through native interfaces, Android and iOS SDKs, and WebAssembly integration. Independent stream contexts share model resources, and consecutive frames from one stream can be processed in a single inference call. We evaluate the released models on EARS-WHAM v2 and compare their quality and processing costs with RNNoise and DPDFNet2. All tested native and Web configurations process files faster than real time, although larger chunks require more time to collect input. We also measure how many streams a shared model can process on a CPU. Our contribution is a configurable deployment system and an experimental study of how model choice, chunk size, and execution environment affect its use. All reported SDK timing uses ONNX Runtime. Index Terms: speech enhancement, streaming inference, cross-platform deployment, temporal frame grouping, computational efficiency
Figures & tables
System
Documented focus
Relation to GigaEdge Spench
Asteroid; ESPnet-SE [ 4 , 5 ]
Research architectures, recipes, and evaluation; ASR integration in ESPnet-SE.
We study the deployment pipeline, host audio integration, and concurrent CPU workloads.
RNNoise [ 1 , 7 , 8 ]
DSP/neural enhancement, pretrained weights and training tools; shared weights with separate states; third-party WebAssembly integration.
Our native efficiency reference, evaluated at 48 kHz and after output downsampling to 8 kHz. We study model selection and temporal frame grouping in configurable channels with different sampling rates.
DTLN; DeepFilterNet [ 2 , 9 ]
Efficient enhancement models with reusable inference and training implementations.
We package model choices together with stream management, platform wrappers, and configurable audio processing.
GTCRN; DPDFNet [ 11 , 12 , 13 ]
Low-cost enhancement architectures; DPDFNet also supplies multirate streaming and deployment formats.
Our experiments address SDK-level chunk costs and shared-model capacity for the released model configurations.
WebRTC APM; SpeexDSP [ 14 , 15 ]
Established integrated audio processing for communication applications.
We combine neural model selection with these types of application components; our AEC reuses SpeexDSP.
sherpa-onnx [ 16 , 17 ]
Cross-platform speech runtime; the inspected upstream streaming denoisers combine a model and one stream state.
We expose model/context separation in the SDK, temporal frame grouping, configurable level control, native/Web AEC, and independent input/output rates. The capacity reference uses a shared-model adaptation of sherpa-onnx.
Table 1: Related software and the focus of this work. The table summarizes documented functions rather than all available features, and does not rank the systems by measured quality or speed.
Figure 1: An example enhancement channel. Stages can be omitted, while normalization can itself combine selected DSP blocks. This is one possible order; available stages and controls depend on the platform interface. The neural stage uses ONNX Runtime or TensorFlow Lite according to model and platform support (Table 2 ).
Model or capability
C/C++
Web
Android
iOS
Spench 8 kHz INT8
ORT
ORT/WASM
—
ORT
Spench 48 kHz Discriminative
ORT, TL
ORT/WASM
TL
ORT
Spench 48 kHz Hybrid
ORT, TL
ORT/WASM
TL
ORT
Neural noise suppression
Yes
Yes
Yes
Yes
Normalization
Yes
Yes
Yes
Yes
Input/output resampling
Yes
Yes
Yes
Yes
Table 2: Platform and model matrix. ORT: ONNX Runtime; TL: TensorFlow Lite. A dash means that the interface does not include the model or expose the function. Backend support and the precision of a released model are separate properties.
Metric
Noisy input
Spench 48 kHz Discriminative
Spench 48 kHz Hybrid
RNNoise 0.2
DPDFNet2 48 kHz HR
ESTOI
0.64
0.76
0.77
0.70
0.81
PESQ-WB
1.23
1.88
1.87
1.53
2.21
SI-SDR (dB)
5.35
12.62
12.86
9.21
14.69
DNSMOS P.808
2.68
3.19
3.45
3.33
3.66
DNSMOS SIG
2.82
3.13
3.33
3.11
3.29
DNSMOS BAK
2.43
3.69
3.84
3.99
4.04
Table 3: Speech quality at 48 kHz on EARS–WHAM v2. Entries are mean scores; all PESQ results use WB mode. Reference-based metrics compensate 16, 20, and 40 ms for Spench, RNNoise, and DPDFNet2, respectively. Bold marks the highest displayed mean in each row.
Metric
Noisy 8 kHz
Spench 8 kHz INT8
RNNoise 48 → 8 kHz
ESTOI
0.64
0.72
0.70
PESQ-NB
1.85
2.28
2.16
SI-SDR (dB)
5.32
10.99
9.03
Table 4: Speech quality of 8 kHz outputs on EARS–WHAM v2. Entries are mean scores; all PESQ results use NB mode. Spench 8 kHz INT8 processes downsampled noisy input; RNNoise 0.2 processes the original 48 kHz input before output downsampling. Bold marks the highest displayed mean; input bandwidths differ.
Model
MMAC/audio s
DNSMOS P.808
SIGMOS OVRL
NISQA MOS
Spench 48 kHz Hybrid
1319.83
3.45
2.73
3.20
RNNoise 0.2
107.33
3.33
2.46
2.17
DPDFNet2 48 kHz HR
2283.97
3.66
2.96
3.81
Table 5: Predicted quality and weighted model arithmetic at 48 kHz. Quality entries are mean scores on EARS–WHAM v2. ONNX counts include forward Mel projections within the graphs; RNNoise counts include retained blocks and recurrent diagonals (dense equivalent: 283.93 MMAC/audio-s). FFT and other external DSP are excluded. Section 5 details the counting rules.
Model
Chunk (ms)
P=1
P=2
P=4
P=8
Spench 8 kHz INT8
16
134.00
297.10
543.90
660.50
64
300.00
581.50
1058.10
1405.60
128
347.80
802.90
1430.40
1772.00
Spench 48 kHz Discriminative
16
17.90
26.70
41.50
49.30
64
34.70
62.20
108.50
142.80
128
49.80
91.40
158.30
212.60
Table 6: Aggregate native throughput (audio s/s).
Configuration
Greatest passing streams
Selected 100Rmaxwall
Spench 8 kHz INT8
593.33±8.33
96.89±2.47
Spench 48 kHz Discriminative
151.00±1.00
98.23±1.39
Spench 48 kHz Hybrid
56.00±1.00
96.71±3.02
sherpa-onnx / DPDFNet2 48 kHz HR a
15.00±0.00
90.12±2.92
RNNoise 0.2
1088.67±30.09
99.58±0.24
Table 7: Shared-model capacity with 16 ms external input chunks and 10 s measurement windows. Values are mean ± sample SD across three searches. Wall RTF is taken from the slowest stream in each search’s selected successful measurement and displayed as 100R (percent of audio duration). For an attempt to pass, every stream’s unrounded wall RTF must be ≤1 .
Model
Path / chunk (ms)
100Rinf
100RSDK
p50
p95
p99
p99inf
Spench 8 kHz INT8
CLI / 16
0.47
0.65
0.07
0.23
0.55
0.37
CLI / 32
0.33
0.52
0.11
0.37
0.82
0.50
CLI / 64
0.23
0.44
0.22
0.57
1.06
0.60
CLI / 128
0.10
0.23
0.26
0.48
0.94
0.40
Web / 16
1.01
1.42
0.19
0.39
0.61
0.42
Web / 32
0.66
1.02
0.30
0.55
0.84
0.54
Table 8: Single-stream processing costs. RTF columns show 100R (percent of audio duration): divide by 100 to recover RTF. These are time ratios, not CPU-use percentages. Call-time percentiles are in milliseconds.
Speech enhancement improves intelligibility and listening quality by suppressing noise, reverberation, and competing speakers in recorded or live audio.