Wenetspeech Wu Bench Eval

wenetspeech-wu-bench-eval

WenetSpeech-Wu: Datasets, Benchmarks, and Models for a Unified Chinese Wu Dialect Speech Processing Ecosystem — Wang et al. (2026) (arXiv:2601.11027, 2026)

What this evaluates

Evaluates speech processing capabilities for the Chinese Wu dialect, including automatic speech recognition (ASR), automatic speech translation (AST), speaker attribute prediction (gender, age), emotion recognition, text-to-speech (TTS), and instruction-following TTS.

Datasets

• WenetSpeech-Wu-Bench — total ?; splits: test (-1); repo https://github.com/ASLP-lab/WenetSpeech-Wu-Repo

Metrics

• CER (%) (primary) — range: percent
  • Character Error Rate, calculated as the number of character substitutions, deletions, and insertions divided by the number of characters in the reference transcript, multiplied by 100.
• SIM — range: [0, 1]
  • Speaker Similarity score, typically derived from embedding cosine similarity or perceptual ratings, measuring how closely the synthesized speech matches the target speaker.
• IMOS/SMOS/AMOS/EMOS/PMOS — range: [1, 5]
  • Mean Opinion Scores for Intelligibility, Speaker MOS, Audio Quality MOS, Emotion MOS, and Prosody MOS. Calculated as the average of subjective Likert-scale ratings (typically 1-5) provided by human listeners.
• AST/Gender/Age/Emotion Accuracy — range: [0, 1]
  • Proportion of correctly predicted labels for translation, gender, age, or emotion categories out of the total number of test instances.

Input / output format

Input: Audio recordings of Wu dialect speech, optionally paired with text transcripts, instruction prompts, or target speaker/emotion tags.

Output: For ASR/AST: transcribed or translated text. For speaker/emotion tasks: predicted categorical labels. For TTS/Instruct TTS: synthesized audio waveform.

Scoring recipe

def compute_cer(predictions, references):
    total_errors = 0
    total_refs = 0
    for pred, ref in zip(predictions, references):
        total_errors += edit_distance(pred, ref)
        total_refs += len(ref)
    return (total_errors / total_refs) * 100

def compute_accuracy(predictions, references):
    correct = sum(1 for p, r in zip(predictions, references) if p == r)
    return correct / len(references)

def compute_mos(ratings):
    return sum(ratings) / len(ratings)

Common pitfalls

• CER is reported as a percentage (e.g., 15.20), not a fraction or decimal.
• TTS evaluation is split into 'easy' and 'hard' subsets based on sample difficulty, requiring separate metric reporting.
• Speaker similarity (SIM) is explicitly excluded for single-speaker fine-tuned models and commercial fixed-speaker systems.

Evidence (verbatim from paper)

Table 4: ASR results (CER%) on various test sets. ... Evaluation is performed on the ASR test set of WenetSpeech-Wu-Bench as well as two in-house manually annotated test sets covering dialogue and reading scenarios, enabling comprehensive assessment across diverse speaking conditions.

Citation

@misc{wang2026wenetspeechwu,
  title={WenetSpeech-Wu: Datasets, Benchmarks, and Models for a Unified Chinese Wu Dialect Speech Processing Ecosystem},
  author={Wang et al. (2026)},
  year={2026},
  note={arXiv:2601.11027}
}

• arXiv: 2601.11027