ONNX Runtime export of AigizK/GigaAM-Bashkir-CV25, a character-level CTC ASR model for Bashkir (Башҡорт теле). The source model was fine-tuned from GigaAM Multilingual multilingual_ssl on Mozilla Common Voice 25 Bashkir.

Results

Raw WER is computed after lowercasing and punctuation removal on the complete Common Voice 25 splits, using onnx-asr with batch size 32.

Runtime	Validation WER	Test WER
PyTorch checkpoint	7.20%	7.73%
ONNX FP32	7.26%	7.74%
ONNX dynamic INT8	7.37%	7.80%

Validation contains 14,525 clips (19.467 hours); test contains 14,572 clips (19.885 hours). See evaluation_results.json for exact error counts, hashes, providers, and runtime measurements.

Installation

CPU:

pip install "onnx-asr[cpu,hub]>=0.12,<0.13"

NVIDIA CUDA for FP32:

pip install "onnx-asr[hub]>=0.12,<0.13" onnxruntime-gpu

Usage

FP32

python

import onnx_asr

model = onnx_asr.load_model("AigizK/GigaAM-Bashkir-CV25-ONNX")
print(model.recognize("bashkir_audio.wav"))

INT8 on CPU

python

import onnx_asr

model = onnx_asr.load_model("AigizK/GigaAM-Bashkir-CV25-ONNX", quantization="int8")
print(model.recognize("bashkir_audio.wav"))

NVIDIA GPU

python

import onnx_asr

model = onnx_asr.load_model(
    "AigizK/GigaAM-Bashkir-CV25-ONNX",
    providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
)
print(model.recognize("bashkir_audio.wav"))

Dynamic INT8 contains ConvInteger/MatMulInteger operators and is intended for CPU. Use FP32 with the CUDA Execution Provider on NVIDIA GPUs.

Batch recognition

Passing a list performs one padded batch and is substantially faster on GPU:

python

texts = model.recognize([
    "audio_01.wav",
    "audio_02.wav",
    "audio_03.wav",
    "audio_04.wav",
])
for text in texts:
    print(text)

Group recordings of similar duration to reduce padding overhead.

Configure CPU threads

python

import onnx_asr
import onnxruntime as ort

options = ort.SessionOptions()
options.intra_op_num_threads = 16
options.inter_op_num_threads = 1

model = onnx_asr.load_model(
    "AigizK/GigaAM-Bashkir-CV25-ONNX",
    quantization="int8",
    providers=["CPUExecutionProvider"],
    sess_options=options,
)

Parallel requests with one shared model

python

from concurrent.futures import ThreadPoolExecutor

files = ["audio_01.wav", "audio_02.wav", "audio_03.wav", "audio_04.wav"]
with ThreadPoolExecutor(max_workers=4) as pool:
    texts = list(pool.map(model.recognize, files))

For maximum GPU throughput, prefer a real batch over concurrent batch-1 calls.

NumPy waveform / non-WAV audio

The built-in path reader accepts mono PCM WAV. Decode MP3/FLAC with an audio library and pass a float32 NumPy array:

python

import soundfile as sf

waveform, sample_rate = sf.read("bashkir_audio.flac", dtype="float32")
text = model.recognize(waveform, sample_rate=sample_rate, channel="mean")
print(text)

Token timestamps and log probabilities

python

timestamped_model = model.with_timestamps()
result = timestamped_model.recognize("bashkir_audio.wav")
print(result.text)
print(result.tokens)
print(result.timestamps)
print(result.logprobs)

Long audio with VAD

python

import onnx_asr

vad = onnx_asr.load_vad("silero")
longform_model = model.with_vad(vad)
for segment in longform_model.recognize("long_bashkir_audio.wav"):
    print(segment.start, segment.end, segment.text)

CLI

onnx-asr AigizK/GigaAM-Bashkir-CV25-ONNX bashkir_audio.wav

onnx-asr accepts mono PCM WAV paths directly. It also accepts a float32 NumPy waveform and a sample_rate argument when audio is decoded by another library. The output is lowercase Bashkir text without punctuation.

Performance benchmark

Higher × realtime is better: 100× means 100 seconds of audio are processed in one second. Audio was decoded to mono 16 kHz and preloaded into RAM. Timed sections include feature extraction, ONNX inference, and greedy CTC decoding; Hub download, model loading/graph optimization, MP3 decoding, and disk I/O are excluded.

Hardware and software:

AMD EPYC 7452, 32 physical cores / 64 threads
NVIDIA GeForce RTX 3090 24 GB
NVIDIA GeForce RTX 5090 32 GB
onnx-asr 0.12.0, ONNX Runtime 1.23.2, Python 3.11
deterministic, evenly spaced rows from Common Voice 25 Bashkir test

Batch size

Same 96 clips / 466.704 seconds of audio for every row. CPU uses 32 ONNX Runtime intra-op threads. GPU uses the CUDA Execution Provider.

Runtime	Batch	Clips/s	Audio speed	p50 batch latency
CPU FP32	1	6.9	33.7×	138.1 ms
CPU FP32	8	6.9	33.6×	1213.8 ms
CPU FP32	32	7.5	36.5×	4103.1 ms
CPU INT8	1	6.3	30.4×	149.2 ms
CPU INT8	8	4.9	23.8×	1659.3 ms
CPU INT8	32	4.8	23.5×	6590.4 ms
RTX 3090 FP32	1	14.3	69.7×	67.9 ms
RTX 3090 FP32	8	110.2	535.7×	74.4 ms
RTX 3090 FP32	32	153.3	745.1×	202.5 ms
RTX 5090 FP32	1	23.8	115.9×	40.3 ms
RTX 5090 FP32	8	178.6	868.3×	46.6 ms
RTX 5090 FP32	32	379.3	1844.0×	86.0 ms

CPU intra-op thread scaling

Batch size 8; same 24 clips / 117.324 seconds of audio.

ORT threads	FP32	INT8
1	5.2×	5.0×
2	9.6×	8.8×
4	17.2×	14.1×
8	26.5×	19.8×
16	34.1×	24.3×
32	38.0×	27.6×

Parallel batch-1 requests

One shared model, Python ThreadPoolExecutor, same 32 clips / 158.580 seconds. CPU sessions use 8 intra-op threads.

Runtime	1 worker	2 workers	4 workers
CPU FP32	29.0×	41.0×	54.4×
CPU INT8	24.9×	40.5×	58.6×
RTX 3090 FP32	76.6×	101.5×	143.8×
RTX 5090 FP32	129.1×	154.1×	203.0×

Key observations:

RTX 5090 reaches about 1844× realtime at batch 32; RTX 3090 reaches about 745×.
Batching is much more effective than concurrent batch-1 calls on GPU.
On this EPYC, FP32 is faster than dynamic INT8. INT8's main benefit here is the 4× smaller artifact and lower memory/storage usage; speed depends on CPU instruction support and ONNX Runtime kernels.
Four parallel CPU requests reach about 54× realtime FP32 and 59× INT8 aggregate throughput.
Results are hardware- and workload-specific. Benchmark your expected audio lengths and concurrency.

Full-precision values and methodology are in benchmark_results.json; the reproducible runner is benchmark_onnx.py.

Artifacts

multilingual_bashkir_ctc.onnx: FP32 encoder and 43-class CTC head.
multilingual_bashkir_ctc.int8.onnx: dynamic unsigned INT8 quantization for CPU.
multilingual_vocab.txt: 42 Bashkir characters plus CTC blank at index 42.
config.json: model configuration used by onnx-asr.
benchmark_results.json: complete CPU/GPU benchmark output.
benchmark_onnx.py: reproducible benchmark runner.

INT8 was produced with ONNX Runtime dynamic quantization, using unsigned 8-bit weights and activations with per-tensor scales. FP32 was evaluated with the CUDA Execution Provider; INT8 was evaluated with the CPU Execution Provider.

Limitations

Trained and evaluated on read speech from Common Voice; conversational and noisy speech may differ.
Greedy CTC decoding only; no external language model.
No capitalization or punctuation prediction.

License

MIT License. Commercial use, modification, redistribution, and private use are permitted. Retain the copyright and license notice when redistributing.