En Fr Translation Bench

en-fr-translation-bench

AI-Driven Modular Services for Accessible Multilingual Education in Immersive Extended Reality Settings: Integrating Speech Processing, Translation, and Sign Language Rendering — Tantaroudas et al. (2026) (arXiv:2604.05591, 2026)

What this evaluates

Evaluates machine translation quality and computational efficiency on a curated set of English-to-French sentences spanning simple, technical, and complex domains. It measures linguistic accuracy alongside inference latency and hardware resource consumption under consumer-grade GPU constraints.

Datasets

• Custom EN-FR Test Set — total 10; splits: test (10)

Metrics

• BLEU score (primary) — range: [0, 100]
  • Standard n-gram overlap metric between reference and hypothesis translations, averaged across the 10 test sentences.
• Avg. Transl. Time (s) — range: seconds
  • Mean inference latency in seconds per translation, measured sequentially to manage limited GPU memory.
• Model Load Time (s) — range: seconds
  • Time required to load the model into GPU VRAM from disk.
• Memory Usage (GB) — range: GB
  • Peak GPU VRAM consumption during inference, measured in gigabytes.
• Successful Translations — range: count
  • Count of translations completed without runtime errors or crashes out of total attempts.

Input / output format

Input: English sentences categorized by complexity: simple conversational phrases, medium-complexity technical sentences, and complex sentences with specialised terminology.

Output: French translation of the input sentence.

Scoring recipe

bleu_scores = []
latencies = []
for sentence in test_dataset:
    start_time = time.time()
    try:
        hypothesis = model.generate(sentence)
        success = True
    except Exception:
        success = False
        continue
    end_time = time.time()
    bleu_scores.append(bleu_score([sentence.ref], hypothesis))
    latencies.append(end_time - start_time)
avg_bleu = sum(bleu_scores) / len(bleu_scores)
avg_latency = sum(latencies) / len(latencies)

Common pitfalls

• Small dataset size (n=10) limits statistical significance and may not generalize across language pairs or domains.
• BLEU scores are calculated on a per-sentence basis rather than corpus-level, which can inflate or deflate scores depending on sentence length and alignment.
• Hardware constraints (RTX 4060, 8GB VRAM) and float16 precision mean latency/memory results are highly specific to this setup and may not transfer to server-grade GPUs.

Evidence (verbatim from paper)

The benchmarking utilised a test dataset of 10 English-to-French translations with varying complexity levels: simple conversational phrases (3 examples), medium-complexity technical sentences (4 examples), and complex sentences with specialised terminology (3 examples). Models were loaded sequentially to manage the limited GPU memory, and inference was conducted using float16 precision to maximise throughput within the available VRAM. For each model, translation quality (BLEU scores), inference speed, and resource utilisation were measured.

Citation

@misc{tantaroudas2026ai,
  title={AI-Driven Modular Services for Accessible Multilingual Education in Immersive Extended Reality Settings: Integrating Speech Processing, Translation, and Sign Language Rendering},
  author={Tantaroudas et al. (2026)},
  year={2026},
  note={arXiv:2604.05591}
}

• arXiv: 2604.05591