OpenSTBench

English | 中文

OpenSTBench is a multidimensional evaluation toolkit for speech translation. It is designed for heterogeneous systems, including speech-to-text translation (S2TT), speech-to-speech translation (S2ST), offline systems, and streaming systems.

The toolkit organizes evaluation into three dimensions:

• Translation Quality: whether the translated text preserves the source meaning.
• Speech Quality: whether generated speech is natural, text-consistent, speaker-preserving, emotion-preserving, and faithful to acoustic events.
• Temporal Quality: whether generated speech preserves duration structure and, for streaming systems, whether output is responsive.

Installation

pip install OpenSTBench

For local development:

git clone https://github.com/sjtuayj/OpenSTBench.git
cd OpenSTBench
conda create -n openstbench python=3.10 -y
conda activate openstbench
pip install -e .

Optional extras:

pip install "OpenSTBench[comet]"
pip install "OpenSTBench[whisper]"
pip install "OpenSTBench[tokenizer-ja]"  # Japanese BLEU
pip install "OpenSTBench[tokenizer-ko]"  # Korean BLEU
pip install "OpenSTBench[speech_quality]"
pip install "OpenSTBench[emotion]"
pip install "OpenSTBench[acoustic-events]"
pip install "OpenSTBench[metricx]"
pip install "OpenSTBench[all]"

MetricX follows the official google-research/metricx runtime requirements. Installing OpenSTBench[metricx] or OpenSTBench[all] pins the MetricX-compatible stack, including transformers[torch]==4.30.2, sentencepiece==0.1.99, datasets==2.13.1, protobuf==3.20.3, and accelerate>=0.26.0.

BLEURT is installed separately:

pip install git+https://github.com/lucadiliello/bleurt-pytorch.git

Package Names

• PyPI package: OpenSTBench
• Python import: openstbench

Evaluation Dimensions

Offline and streaming are supported system settings, not separate metric dimensions. Use the evaluators that match the available outputs: text, generated speech, source/target audio pairs, event annotations, or streaming traces.

Acoustic-event evaluation

Acoustic-event evaluation measures whether the target audio preserves the annotated event categories from the source audio. It does not evaluate when an event occurs. OpenSTBench runs the published BEATs_strong_1 checkpoint without training or fine-tuning and computes threshold-free, clip-based macro Average Precision (event_content_map). Classes enter the macro average only when the evaluated cohort contains both positive and negative clips.

The public AcousticEventEvaluator and compute_event_content_map APIs return a Content mAP result directly, including the overall event_content_map, per-class AP, positive/negative counts, and class coverage. See examples/python/acoustic_event_eval.py for a runnable example.

The newtest workflow writes metrics_acoustic_events.json with two metric views and separate coverage:

• metrics.successful_only.event_content_map is the official score over successfully generated target audio.
• metrics.all_samples_zero_scored.event_content_map is diagnostic only; failed target audio receives zero confidence for every event class before mAP is recomputed on the complete cohort.
• coverage reports the success rate, failure rate, and failure reasons.

Omit --beats_model_path in newtest/eval/eval_acoustic_events.py to download the fixed checkpoint into the OpenSTBench user cache, or provide an explicit checkpoint path. There is no training command or manual detection threshold.

Experimental Overview

The radar plot below illustrates the multidimensional view produced by OpenSTBench for representative streaming and offline speech translation systems. It summarizes how systems can differ across translation quality, speech quality, and temporal quality: a system with strong translation quality may still show different behavior in speech realization, speaker or emotion preservation, acoustic-event fidelity, temporal consistency, and latency or efficiency.

Datasets

The paper uses the following datasets. Please follow the license and access terms of each original dataset.

Quick Start

from openstbench import TranslationEvaluator

evaluator = TranslationEvaluator(
    use_bleu=True,
    use_chrf=True,
    use_comet=False,
    use_bleurt=False,
    use_metricx=True,
    device="cuda",
)

scores = evaluator.evaluate_all(
    reference=["我喜欢看电影。", "今天天气很好。"],
    target_text=["我喜欢看电影。", "今天天气很好。"],
    source=["I like watching movies.", "The weather is nice today."],
    target_lang="zh",
    asr_text=["我喜欢看电影。", "今天天气很好。"],
)

print(scores)

ASR translation-quality evaluation is enabled by default and runs only when target_audio or precomputed asr_text is supplied. To transcribe generated speech directly, pass a file, a list of files, or a directory as target_audio; Whisper defaults to medium and can be changed with TranslationEvaluator(whisper_model="large-v3"). Set use_asr=False to disable every ASR_ metric.

Route languages to different ASR models when needed:

from openstbench import ASRRouter, TranslationEvaluator, WhisperASRBackend

router = ASRRouter({
    "default": WhisperASRBackend(model="medium"),
    "ja": WhisperASRBackend(model="large-v3"),
})
evaluator = TranslationEvaluator(asr_router=router)

Keep Cantonese as the independent route "yue"; if the selected Whisper checkpoint does not support it, configure a Cantonese ASRBackend or provide precomputed asr_text. Dzongkha ("dz") likewise uses character BLEU/CER but requires a custom ASR backend or asr_text with the default Whisper backend. For language-aware latency tokenization, use --target-language ja --latency-unit auto; omitting auto preserves the existing explicit unit behavior.

Multilingual quick start

target_lang is the language switch for translation quality and speech consistency. Install the optional tokenizer before evaluating Japanese or Korean BLEU:

pip install "OpenSTBench[tokenizer-ja]"  # Japanese
pip install "OpenSTBench[tokenizer-ko]"  # Korean

scores = evaluator.evaluate_all(
    reference=["今日は天気がとても良いです。"],
    target_text=["今日は天気がとても良いです。"],
    asr_text=["今日は天気がとても良いです。"],
    target_lang="ja",
)

For streaming latency, pass the target language and let auto choose the unit:

python -m openstbench.latency.cli \
  --source data/source.txt --target data/ref.txt \
  --task s2t --agent-script my_agent.py --agent-class MyAgent \
  --target-language ja --latency-unit auto

See translation_eval.py, speech_quality_eval.py, and latency_eval.py for complete parameter templates. Cantonese and other languages unsupported by a selected Whisper checkpoint require a custom ASRBackend or precomputed asr_text.

Supported languages

OpenSTBench's built-in language policy is defined in src/openstbench/language_policy.py. target_lang accepts normalized ISO-style language codes and common aliases, including eng -> en, cmn/zho/chi -> zh, jpn -> ja, and kor -> ko.

OpenSTBench does not reject language codes that are not listed below. For unlisted languages, translation metrics use the default space-delimited policy: SacreBLEU 13a, WER_Consistency, and word-level latency tokenization. ASR availability is backend-dependent: use a compatible WhisperASRBackend, a custom ASRBackend, or precomputed asr_text.

Examples

Complete parameter templates are kept in examples/. The README intentionally stays compact; use these files for configurable parameters, input formats, and output fields.

• examples/python/translation_eval.py
• examples/python/speech_quality_eval.py
• examples/python/speaker_similarity_eval.py
• examples/python/emotion_eval.py
• examples/python/acoustic_event_eval.py
• examples/python/temporal_consistency_eval.py
• examples/python/latency_eval.py
• examples/bash/install_extras.sh
• examples/bash/run_latency_cli.sh

Latency can also be run from the module CLI:

python -m openstbench.latency.cli --help

Conventions

• Text inputs generally accept list[str], one-sample-per-line .txt files, and .json files where supported by the evaluator.
• Audio inputs generally accept folders, list[str], .txt path lists, and .json path lists where supported by the evaluator.
• Built-in language tokenization, speech-consistency, and latency_unit="auto" behavior are summarized in Supported languages.
• ASRRouter can route target languages to different WhisperASRBackend checkpoints or custom ASRBackend implementations. If a Whisper checkpoint does not expose the requested language, ASR returns an unavailable/empty transcript instead of passing an invalid language token to the model.
• Latency tokenization accepts unit="auto" (CLI: --latency-unit auto) together with a target language. Existing explicit word and char behavior is unchanged.
• Evaluators that accept pretrained model sources use a local-first rule. If the supplied local path exists, OpenSTBench uses it; otherwise it falls back to the configured remote model id.
• Optional dependencies are loaded only when the corresponding evaluator needs them.
• MetricX is enabled by default in TranslationEvaluator. It follows the official google-research/metricx README, uses text only, reports error scores in [0, 25] where lower is better, and can be disabled with use_metricx=False.
• TranslationEvaluator also enables ASR variants by default. With target_audio, Whisper transcribes the generated speech before the same enabled translation metrics are applied; with asr_text, the supplied transcript is used directly. Calls without either input preserve the text-only behavior.

Acknowledgements

• We especially thank SimulEval, from which parts of OpenSTBench's latency evaluation components are adapted
• sacreBLEU, COMET, MetricX, and bleurt-pytorch, a PyTorch port of BLEURT, for translation quality evaluation
• Whisper, SpeechMOS/UTMOS, Resemblyzer, and WavLM for speech quality and speaker similarity evaluation
• FunASR and Emotion2Vec for emotion preservation evaluation
• PretrainedSED, BEATs, and sed-scores-eval for acoustic-event evaluation

Citation

If you find our work useful, please cite as：

@misc{an2026openstbenchsemanticevaluationspeech,
      title={OpenSTBench: Beyond Semantic Evaluation for Speech Translation}, 
      author={Yanjie An and Yuxiang Zhao and Yichi Zhang and Qixi Zheng and Yujie Tu and Keqi Deng and Kai Yu and Xie Chen},
      year={2026},
      eprint={2605.30792},
      archivePrefix={arXiv},
      primaryClass={eess.AS},
      url={https://arxiv.org/abs/2605.30792}, 
}

License

OpenSTBench's original code is released under the MIT License. See LICENSE.

Some latency evaluation components include code adapted from SimulEval, which is licensed under the Creative Commons Attribution-ShareAlike 4.0 International License (CC BY-SA 4.0). Those adapted portions are distributed under CC BY-SA 4.0. See THIRD_PARTY_NOTICES.md for details.

The datasets referenced by OpenSTBench, including the datasets used in the paper, are not covered by the OpenSTBench code license. They are provided by their original authors or distributors under their own licenses and terms of use. Some datasets are restricted to research or non-commercial use.