bc7ec356/heep-universal

Post-Rebuttal Update: Cross-Architecture Validation with HEEP-Indic

Addressing Q1 (Gain Attribution), Q2 (Baselines), and Q3 (Base Model Dependency)

We apologize for the supplementary post after the rebuttal period. These results were finalized shortly after the deadline, and we wanted to ensure complete experimental evidence was available rather than leave placeholders.

🔗 Resources

• Reproducibility (Universal Model): https://huggingface.co/bc7ec356/heep-universal
• Cross-Architecture Model (Indic): https://huggingface.co/bc7ec356/heep-indic

Cross-Architecture Generalization

To directly address concerns about generalization beyond Whisper V3 Turbo, we trained Qwen3-ASR (1.7B), an architecturally distinct audio-language model, on HEEP-curated data spanning 46 Indian languages (~4.78M utterances). The curation pipeline is identical to the one described in the paper with no architecture-specific tuning.

Hindi Benchmark Comparison (7 Benchmarks)

Model	Kathbath	Kathbath Noisy	CommonVoice	FLEURS	IndicTTS	RESPIN	Gramvaani	Avg
Google STT	14.3	16.7	20.8	19.4	18.3	–	59.9	24.9
IndicWav2Vec	12.2	16.2	20.2	18.3	15.0	–	42.1	20.7
Azure STT	13.6	15.1	14.6	24.3	15.2	–	42.3	20.8
Nvidia Conformer-CTC Large	12.7	14.2	21.2	15.7	12.2	–	42.6	19.8
IndicWhisper	10.3	12.0	15.0	11.4	7.6	–	26.8	13.8
HEEP-Indic	8.53	8.97	9.96	11.04	6.59	12.05	25.98	11.9

HEEP-Indic achieves 11.9% average Hindi WER vs. 13.8% for IndicWhisper (14% relative improvement).

Multilingual Results (16 Languages)

Dataset	Ben	Bho	Chh	Guj	Hin	Kan	Mag	Mai	Mal	Mar	Odi	Pun	San	Tam	Tel	Urd	Avg
Kathbath	14.6	–	–	17.4	8.5	23.0	–	–	39.3	19.2	25.4	15.8	41.4	30.3	29.0	12.1	23.0
Kathbath Hard	15.7	–	–	18.5	9.0	25.1	–	–	41.2	20.4	27.7	16.6	43.6	32.6	30.3	11.9	24.4
CommonVoice	21.0	–	–	–	10.0	–	–	–	46.0	21.5	34.6	17.5	–	34.0	–	20.6	25.7
FLEURS	22.4	–	–	23.3	11.0	23.1	–	–	34.4	25.5	33.3	25.0	–	35.1	31.9	22.4	26.1
IndicTTS	15.8	–	–	16.9	6.6	19.6	–	–	26.4	14.5	14.8	–	–	22.6	31.3	–	18.7
Gramvaani	–	–	–	–	26.0	–	–	–	–	–	–	–	–	–	–	–	26.0
RESPIN	32.5	21.3	21.6	–	12.1	45.6	27.7	41.1	–	32.7	–	–	–	–	37.5	–	30.2
Avg	20.4	21.3	21.6	19.0	11.9	27.3	27.7	41.1	37.5	22.3	27.2	18.7	42.5	30.9	32.0	16.7	24.6

Key Takeaways

1. Cross-architecture generalization confirmed. The same HEEP pipeline improves two distinct backbones: Whisper V3 Turbo (0.8B, encoder-decoder) and Qwen3-ASR (1.7B, audio-language model), without modification.

2. Controlled multilingual evaluation. Results span 16 languages across Indo-Aryan, Dravidian, and Classical families on standardized benchmarks with consistent evaluation protocols.

3. Model-independent scoring. Entropy scoring operates on MFCCs, G2P phonemes, and token distributions, not model internals. The same curated dataset was used for both backbones.

4. Reproducibility. Model weights, curation code, and training scripts for both backbones are at the anonymous repository.

We hope Reviewers 2ezj, oXjG, and S4Jd also find this supplementary evidence relevant to their earlier questions on generalization and controlled multilingual evaluation.

---

Model Overview

HEEP Universal supports transcription across 204 languages, including a wide range of Indic and global languages, with consistent performance across various domains such as meetings, earnings calls, broadcast media, and educational content. The model is optimized for high-precision, verbatim transcription capturing spoken content word-for-word with remarkable fidelity.

Core Insight: Strategic selection of high-entropy samples leads to better ASR models than training on larger but redundant datasets.

HEEP Methodology

HEEP (High Entropy Exponential Pruning) is an entropy-based data curation methodology that prioritizes information density over data quantity. It identifies high-information training samples while progressively filtering redundant data, enabling efficient model training with significantly reduced computational resources.

Mathematical Foundation

Sample Score (Equation 1)

The information score for each sample combines multiple entropy dimensions:

S(x) = α₁·H_acoustic(x) + α₂·H_phonetic(x) + α₃·H_linguistic(x) + α₄·H_contextual(x) + β·MI(x, D)

Where:

• H_acoustic(x): Spectral/MFCC entropy measuring acoustic diversity
• H_phonetic(x): Phoneme distribution entropy capturing phonetic complexity
• H_linguistic(x): Vocabulary and syntax entropy measuring linguistic richness
• H_contextual(x): Domain and discourse entropy
• MI(x, D): Mutual information contribution relative to dataset
• α₁...α₄, β: Configurable weights (default: 0.25, 0.20, 0.25, 0.15, 0.15)

Mutual Information (Equation 2)

The mutual information between acoustic features and transcription:

I(x, y) = Σ_{j,ℓ} p(f_j, y_ℓ) log [p(f_j, y_ℓ) / (p(f_j)·p(y_ℓ))]

Selection Criterion

Samples are selected based on a threshold:

D' = {x ∈ D : S(x) > τ}

Progressive Filtering (Equation 8)

The threshold increases exponentially across rounds:

τ_{k+1} = τ_k · growth_factor

Error-Aware Adaptation

After each training round, sample scores are adjusted based on model errors:

S'(x) = S(x) + λ_err·ErrorRelevance(x, errors_k) + λ_cross·CrossLingualOverlap(x)

Algorithm Overview

Algorithm: HEEP Data Curation with Error-Aware Adaptation

Input: Dataset D, initial threshold τ₀, growth factor g
Output: Curated dataset D*

1. Initialize scorer with entropy estimators
2. Fit scorer to D (compute normalization stats, fit MI estimator)
3. D* ← D
4. k ← 0
5. While |D*| > min_samples AND k < max_rounds:
    a. For each x in D*:
        Compute S(x) = Σᵢ αᵢ·Hᵢ(x) + β·MI(x, D)
    b. If error_patterns available:
        Adjust S'(x) = S(x) + λ_err·ErrorRelevance(x) + λ_cross·CrossLingualOverlap(x)
    c. D* ← {x ∈ D* : S'(x) > τₖ}
    d. If train_callback: Train model on D*
    e. If eval_callback: Analyze errors, update error_patterns
    f. τₖ₊₁ ← τₖ · g
    g. k ← k + 1
6. Return D*

Key Benefits

• Training on 10-20% of data while matching or exceeding full-dataset performance
• Efficient multilingual model development with cross-lingual transfer
• Error-aware adaptive sample selection across training rounds
• Significant reduction in computational resources and training time

Performance Benchmarks

OpenASR Leaderboard Results

Dataset	WER (%)	RTFx
AMI Test	4.19	70.22
Earnings22 Test	5.83	101.52
GigaSpeech Test	4.99	131.09
LibriSpeech Test Clean	0.71	158.74
LibriSpeech Test Other	2.17	142.40
SPGISpeech Test	1.10	170.85
TedLium Test	1.43	153.34
VoxPopuli Test	4.34	179.28

Composite Results

• Overall WER: 3.10%
• Average RTFx: 146.23

RTFx (Real-Time Factor) indicates inference speed relative to audio duration. Higher values mean faster processing.

Model Details

• Architecture: Transformer-based encoder-decoder optimized for multilingual transcription
• Languages: 204 languages supported
• Format: Transformers compatible (safetensors)
• Sampling Rate: 16 kHz
• Precision: FP16/FP32 supported
• Optimization: Real-time inference capable with GPU acceleration

Key Features

• Exceptional Accuracy: Achieves 3.10% WER across diverse English test sets
• Real-Time Performance: Average RTFx of 146.23 enables real-time applications
• Verbatim Transcription: Optimized for accurate, word-for-word transcription
• Multi-Domain Excellence: Superior performance across conversational, broadcast, and read speech
• Multilingual Support: 204 languages with cross-lingual transfer learning
• HEEP-Curated Training: Strategic entropy-based data selection for maximum information density

Usage

from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
import torch

device = "cuda:0" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

model = AutoModelForSpeechSeq2Seq.from_pretrained(
    "bc7ec356/heep-universal",
    torch_dtype=torch_dtype,
    use_safetensors=True,
)
model.to(device)

processor = AutoProcessor.from_pretrained("bc7ec356/heep-universal")

pipe = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    torch_dtype=torch_dtype,
    device=device,
)

result = pipe("audio.wav")
print(result["text"])

Use Cases

HEEP Universal excels in various speech recognition scenarios:

• Meeting Transcription: High accuracy on conversational speech (AMI: 4.19% WER)
• Financial Communications: Specialized performance on earnings calls (Earnings22: 5.83% WER)
• Broadcast Media: Excellent results on news, podcasts, and media content
• Educational Content: Optimized for lectures and presentations
• Customer Support: Accurate transcription of support calls
• Legal Documentation: Professional-grade accuracy for legal proceedings
• Medical Transcription: High-quality transcription for medical consultations

Performance Optimization Tips

• GPU Acceleration: Use device="cuda" for significantly faster inference
• Precision: Set torch_dtype=torch.float16 for optimal speed on modern GPUs
• Language Specification: Specify language code when known to improve accuracy and speed
• Beam Size: Use beam_size=5 for best accuracy, reduce for faster inference
• Batch Processing: Process multiple files with a single model instance for efficiency

Acknowledgments

HEEP Universal was developed using the HEEP framework for entropy-based data curation. We thank the open-source community for providing foundational tools that make this work possible.

Citation

If you use this model in your research, please cite:

@article{anonymous2026heep,
  title={HEEP: High Entropy Exponential Pruning for State-of-the-Art ASR Through Strategic Data Curation},
  author={Anonymous},
  journal={Under Review},
  year={2026}
}