nur-dev/multilingual-asr

FastConformer Multilingual ASR

A multilingual automatic speech recognition model supporting Kazakh, Russian, Uzbek, and English. Built on NVIDIA NeMo's FastConformer-CTC architecture.

Languages

Language	Code
Kazakh	kk
Russian	ru
Uzbek (Latin)	uz
English	en

Results

Full test set (76,739 samples):

Language	Samples	CER	WER
Russian	10,203	2.34%	9.16%
Kazakh	33,964	8.27%	14.09%
Uzbek	16,184	7.10%	28.82%
English	16,388	9.53%	22.29%
Overall	76,739	7.73%	16.86%

Usage

Requirements

pip install nemo_toolkit[asr]

Transcribe Audio Files

import nemo.collections.asr as nemo_asr

model = nemo_asr.models.ASRModel.restore_from("fastconformer_multilingual.nemo")
model.freeze()

transcriptions = model.transcribe(["audio.wav"])
print(transcriptions[0])

Real-Time Streaming Transcription

import nemo.collections.asr as nemo_asr
import sounddevice as sd
import numpy as np
import queue
import tempfile
import os
import soundfile as sf

SAMPLE_RATE = 16000
CHUNK_SEC = 3
CHUNK_SAMPLES = SAMPLE_RATE * CHUNK_SEC

model = nemo_asr.models.ASRModel.restore_from("fastconformer_multilingual.nemo")
model.freeze()

audio_queue = queue.Queue()


def audio_callback(indata, frames, time, status):
    audio_queue.put(indata[:, 0].copy())


def transcribe_stream():
    buffer = np.array([], dtype=np.float32)
    with sd.InputStream(
        samplerate=SAMPLE_RATE,
        channels=1,
        callback=audio_callback,
        blocksize=SAMPLE_RATE,
    ):
        print("Listening... (Ctrl+C to stop)")
        while True:
            chunk = audio_queue.get()
            buffer = np.concatenate([buffer, chunk])
            if len(buffer) >= CHUNK_SAMPLES:
                tmp = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
                sf.write(tmp.name, buffer[:CHUNK_SAMPLES], SAMPLE_RATE)
                result = model.transcribe([tmp.name])
                text = result[0] if isinstance(result[0], str) else result[0].text
                if text.strip():
                    print(f"> {text}")
                os.unlink(tmp.name)
                buffer = buffer[CHUNK_SAMPLES:]


if __name__ == "__main__":
    transcribe_stream()

Batch Transcription

import nemo.collections.asr as nemo_asr
from pathlib import Path

model = nemo_asr.models.ASRModel.restore_from("fastconformer_multilingual.nemo")
model.freeze()

audio_files = list(Path("audio_dir").glob("*.wav"))
transcriptions = model.transcribe([str(f) for f in audio_files], batch_size=32)

for path, text in zip(audio_files, transcriptions):
    t = text if isinstance(text, str) else text.text
    print(f"{path.name}: {t}")

Model Details

• Architecture: FastConformer-CTC
• Framework: NVIDIA NeMo
• Audio: 16kHz mono WAV

Limitations

• Optimized for clear speech; performance may degrade on noisy audio
• No punctuation or capitalization in output
• Language is auto-detected, not explicitly specified

License

This model is released under CC BY-NC 4.0. Non-commercial use only.

Evaluation (independently measured)

Held-out public test sets, measured directly — not self-reported (seed 42, uniform multilingual-Whisper normalization). FLEURS test = 500 utterances/language; ISSAI KSC2 test = 1000 utterances (in-domain Kazakh, spanning crowd/parliament/podcasts/radio/talkshow).

Test set	Lang	WER (%)	CER (%)
FLEURS kk_kz	kk	16.96	10.29
FLEURS ru_ru	ru	17.16	7.61
FLEURS en_us	en	18.34	8.04
ISSAI KSC2	kk	33.10	22.81

Macro WER (kk/ru/en): 17.49% (unweighted mean; penalises models that do not cover all three languages).

Note. The card reports all 16.86 (author eval set); independently measured macro is 17.49 — the same ballpark on a different, held-out set.

License & commercial use

Non-commercial use only (CC BY-NC 4.0). For commercial licensing or other inquiries, please reach out to the author, Nurgali Kadyrbek, on LinkedIn: https://www.linkedin.com/in/nurgali-kadyrbek-504260231/