✅ This is the recommended variant

This is the general-purpose Hinglish model of the series. It is trained on the union of both datasets and improves over the base Qwen3-ASR-0.6B on BOTH domains at once — conversational speech and technical-tutorial speech (see results below).

Unlike the v2 OpenSLR specialist, v3 does not regress on conversational speech: it bridges the two domains instead of trading one off against the other. If you want a single model that just works across conversational and tutorial Hinglish, use this one.

If you instead want to squeeze out the last 1-2 pp in a single, known domain, the domain-specialists are still available:

• Surajgameramp/qwen3-asr-0.6b-hinglish-hiacc-v1 — conversational (HiACC) specialist.
• Surajgameramp/qwen3-asr-0.6b-hinglish-openslr104-v2 — technical-tutorial (OpenSLR-104) specialist (note: regresses on conversational speech).

Result matrix

v3 is the only variant that beats the base model on both test sets. It sacrifices only ~1.6 pp vs the HiACC specialist on conversational speech and ~2.2 pp vs the OpenSLR specialist on tutorial speech, in exchange for being usable across both domains.

HiACC cohort split for v3: adult 15.41%, children 16.66%.

WER computed with jiwer after lowercase + punctuation-stripping normalization applied symmetrically to predictions and references.

Intended use

• Transcribing both conversational Hindi-English code-switched speech (casual Q&A, storytelling, image-prompted descriptions) and technical-tutorial Hinglish (software walkthroughs, lecture-style instruction).
• Mixed Devanagari + Latin script output is the natural format — the model emits Devanagari for Hindi words and Latin for English words, matching how Hinglish is typically written.

Usage

python
Copy code
import torch
from qwen_asr import Qwen3ASRModel

model = Qwen3ASRModel.from_pretrained(
    "Surajgameramp/qwen3-asr-0.6b-hinglish-union-v3",
    dtype=torch.bfloat16,
    device_map="cuda:0",
    attn_implementation="flash_attention_2",
)

results = model.transcribe(audio="path/to/your.wav", language=None)
print(results[0].text)
# Example: "मेरा favourite festival Diwali है"

Inference requires the qwen-asr package (pip install qwen-asr==0.0.6).

Training data

The union of two Hindi-English code-switched corpora, simply concatenated with no upsampling:

• HiACC (Singh, Singh & Kadyan, 2025): 5.24 h of conversational Hinglish, 16 kHz mono WAV. CC BY 4.0.
• OpenSLR-104 Hindi-English (released for the MUCS-2021 Multilingual & Code-Switching ASR challenge): 89.86 h of Hindi-English code-switched spoken-tutorial speech from IIT Bombay's Spoken Tutorial project. CC BY 4.0.

Combined splits:

• Train 53,627 utterances (HiACC 6.8% + OpenSLR-104 93.2%)
• Val 3,282 utterances

Each corpus's own test set is used for evaluation, reported separately in the result matrix above.

Training procedure

• Base model: Qwen/Qwen3-ASR-0.6B
• Script: qwen3_asr_sft.py from QwenLM/Qwen3-ASR at commit c17a131f
• Fine-tune scope: full-parameter (no frozen layers, no LoRA — all AuT encoder + projector + LLM weights updated)
• Hardware: 2× NVIDIA H100 80GB on Modal
• Optimizer: AdamW, LR 2e-5, linear schedule, warmup_ratio 0.02
• Effective batch: 32 (per-device 8 × grad-accum 2 × 2 GPUs)
• Precision: bf16 + FlashAttention 2
• Epochs: 2 (3,352 total steps); best checkpoint by eval_loss was step 3200 (epoch 1.91, eval_loss=0.1500)
• Wall-clock: 2,943 s (~49 min)

Data format

Each training example uses language None<asr_text>... as the target prefix (language-agnostic decoding, following Polyglot-Lion which adapts Toshniwal et al., 2018). Transcripts are kept in their natural mixed-script form (Devanagari for Hindi, Latin for English).

Limitations

• Generalist trade-off: slightly worse than each specialist in its own domain (~1.6 pp behind v1 on HiACC, ~2.2 pp behind v2 on OpenSLR-104).
• OpenSLR-104 / tutorial test WER (35.06%) is still substantial. Tutorial speech with dense code/path/version vocabulary remains harder than conversational data.
• Single-seed, single-config training. No hyperparameter sweep.
• WERs not directly comparable to MUCS-2021 published baselines without matching their Kaldi-style normalization conventions.

Citation

If you use this model, please cite the underlying works:

bibtex
Copy code
@article{shi2026qwen3asr,
  title   = {Qwen3-ASR Technical Report},
  author  = {Shi, Xian and Wang, Xiong and others},
  year    = {2026},
  url     = {https://arxiv.org/abs/2601.21337}
}

@article{dang2026polyglot,
  title   = {Polyglot-Lion: Efficient Multilingual ASR for Singapore via
             Balanced Fine-Tuning of Qwen3-ASR},
  author  = {Dang, Quy-Anh and Ngo, Chris},
  year    = {2026},
  url     = {https://arxiv.org/abs/2603.16184}
}

@misc{singh2025hiacc,
  title  = {HiACC: Hinglish Adult \& Children Code-switched Corpus},
  author = {Singh, Shruti and Singh, Muskaan and Kadyan, Virender},
  year   = {2025},
  doi    = {10.5281/zenodo.15551669},
  url    = {https://zenodo.org/records/15551669}
}

@inproceedings{diwan2021mucs,
  title     = {{MUCS} 2021: Multilingual and Code-Switching {ASR} Challenges for Low Resource Indian Languages},
  author    = {Diwan, Anuj and Vaideeswaran, Rakesh and Shah, Sanket and others},
  booktitle = {Proc. Interspeech 2021},
  year      = {2021}
}

License

Apache 2.0, inherited from the base Qwen3-ASR-0.6B model. HiACC training data is CC BY 4.0; OpenSLR-104 corpus is CC BY 4.0 (see openslr.org/104 for full license terms).