openarcane/openarcane-asr-0.6b-preview

OpenArcane ASR 0.6B — Experimental Preview

OpenArcane ASR is a speech recognition model for Indian languages. This first release covers Kannada, Malayalam, Tamil and Telugu.

Experimental preview. This is an early release for testing and feedback. Expect errors, and do not use it in production yet.

Model details

Base model	Qwen/Qwen3-ASR-0.6B
Fine-tuning	LoRA adapter, with an extended tokenizer for the four scripts
Languages	Kannada, Malayalam, Tamil, Telugu
Other languages	Hindi, English, Chinese and Japanese from the base model still work, but are not the focus
Output	Text in the native script
License	Apache-2.0

Usage

pip install qwen-asr==0.0.6 peft==0.17.1 soundfile

from huggingface_hub import hf_hub_download
import importlib.util

path = hf_hub_download("openarcane/openarcane-asr-0.6b-preview", "inference.py")
spec = importlib.util.spec_from_file_location("openarcane_asr", path)
oa = importlib.util.module_from_spec(spec); spec.loader.exec_module(oa)

model = oa.load()                                      # GPU if available, else CPU
print(oa.transcribe(model, "clip.wav", "Kannada"))     # language given (recommended)
print(oa.transcribe(model, "clip.wav"))                # automatic language detection

Audio can be a file path or a (numpy_array, sample_rate) tuple. Command line: python inference.py clip.wav --language Telugu.

Examples

Selected clips from the held-out FLEURS test and IndicVoices validation sets where the model output matches the reference transcript word for word. These are hand-picked, not typical output; see Limitations.

Audio	Model output (language given)
Kannada · FLEURS test · 8.3 s	ಶಿಲ್ಪಗಳ ಮೇಲೆ ಗುರುತುಗಳನ್ನು ಮಾಡುವ ಮೂಲಕ ಅಥವಾ ಗ್ರಾಫಿಟಿ ಬರೆಯುವ ಮೂಲಕ ಅದಕ್ಕೆ ಧಕ್ಕೆ ತರಬೇಡಿ
Kannada · FLEURS test · 10.3 s	ಕೆಲವು ಪ್ರದೇಶಗಳಲ್ಲಿ ಒಂದು ನಿಮಿಷ ಕುದಿಯುವ ನೀರು ಸಾಕು. ಇತರ ಪ್ರದೇಶಗಳಲ್ಲಿ ಇದು ಹಲವಾರು ನಿಮಿಷಗಳನ್ನು ತೆಗೆದುಕೊಳ್ಳುತ್ತದೆ.
Malayalam · FLEURS test · 11.2 s	അവന് വേദനയുണ്ടായിട്ടും ഗെയിംസിൽ നിരോധിക്കപ്പെട്ടതിനാൽ മരുന്നുകൾ കഴിക്കാൻ സാധിച്ചില്ല
Tamil · IndicVoices valid · 8.7 s	மணிப்பூர் மாநிலத்தில் கார்பவேக்ஸ் அளிக்கும் அனைத்து மையங்களையும் காட்டவும்
Telugu · FLEURS test · 5.4 s	ఆ విధంగా పెన్సిల్ బయటకు వచ్చినప్పుడు చాలా మందికి మంచి స్నేహితుడు

Limitations

• Punctuation is partial: mostly in Malayalam. Kannada, Tamil and Telugu output is mostly unpunctuated.
• Numbers are written as words (for example, "1200" comes out as "ಸಾವಿರದ ಇನ್ನೂರು").
• Conversational and noisy speech is harder than read speech.
• Give the language when you know it. Automatic detection is less reliable, especially for Telugu and very short clips.

Roadmap

This preview is the first step. Planned for upcoming releases:

• Support for more Indian languages.
• Higher accuracy, especially on conversational and noisy speech.
• Consistent punctuation across all supported languages.
• Numbers written as digits.

Credits

Trained on IndicVoices, Kathbath, SPRING-INX, FLEURS, LibriTTS-R, IndicTTS English, AISHELL-1 and MUSAN. The example clips are from FLEURS and IndicVoices (CC-BY-4.0).

IndicTTS data: COPYRIGHT 2016 TTS Consortium, TDIL, MeitY represented by Hema A Murthy & S Umesh, Department of Computer Science and Engineering and Electrical Engineering, IIT Madras.

License

Apache-2.0, the same as the base model.