This task focuses on the real time (also known as simultaneous or streaming) aspect of speech and machine translation as it enables interesting applications such as simultaneous interpretation or international conference live translations. Real-time systems are typically evaluated with respect to quality and latency. This year, we will have 3 tracks:

  • Text-to-Text: translating ground-truth transcripts in real-time from English to German, English to Japanese.
  • Speech-to-Text: directly translating speech into text in real-time from English to German.

We encourage participants to enter all tracks when possible. We also encourage participants to contrast cascaded and end-to-end solutions for the Speech-to-Text track.


We will use a very similar system as last year for evaluation. The system’s performance will be evaluated in two ways:

  • Translation quality: we will use multiple standard metrics: BLEU, TER, and METEOR.
  • Translation latency: we will make use of the recently developed metrics for simultaneous machine translation including average proportion (AP), average lagging (AL) and differentiable average lagging (DAL).

This year, the evaluation implementation will use the SimulEval toolkit. For latency measurement, we will contrast computation aware and non computation aware latency metrics. See the SimulEval description for how those metrics are defined. Note that the definition of average lagging has been modified from the original definition (see section 3.2 in the SimulEval description). Finally, we will run systems on the segmented and unsegmented blind test set. Note the latter setting requires participants to have segmentation logic in their systems. The latency is calculated on word level for En-De systems and character level for En-Ja systems.

The participants will submit a Docker image (see below for an example) and the organizers will run the image in a controlled environment, specifically an ap3.2xlarge AWS instance (see details in


We will evaluate translation quality with detokenized BLEU and latency with AP, AL and DAL. The systems will be ranked by the translation quality with different latency regimes. Three regimes, low, medium and high, will be evaluated. Each regime is determined by a maximum latency threshold. The thresholds are determined by AL, which represents the delay to the perfect real time system (milliseconds for speech and number of words for text), but all three latency metrics, AL, DAL and AP will be reported. Based on analysis on the quality-latency tradeoffs for the baseline systems, the thresholds are set as follows:

Speech Translation:

  • Low latency: AL < = 1000
  • Medium latency: AL < = 2000
  • High Latency: AL < = 4000

Text Translation (English-German)

  • Low latency: AL < = 3
  • Medium Latency: AL < = 6
  • High Latency: AL < = 15

Text Translation (English-Japanese)

  • Low latency: AL < = 8
  • Medium Latency: AL < = 12
  • High Latency: AL < = 16

The submitted systems will be categorized into different regimes based on the AL calculated on the Must-C English-German test set (tst-COMMON) for English-German or on the IWSLT21 dev set for English-Japanese, while the translation quality will be calculated on the blind test set. We require participants to submit at least one system for each latency regime. Participants are encouraged to submit multiple systems for each regime in order to provide more data points for latency-quality tradeoff analyses. If multiple systems are submitted, we will keep the one with the best translation quality for ranking. In addition, within each latency regime, we will also measure computation aware AL and rank systems accordingly. Finally, we will report latency-quality trade-off curves for non computation aware AL and for computation aware AL in the findings paper.

Training and Development Data

Text-to-Text and Speech-to-Text Tracks (English to German)

You may use the same training and development data available for the Offline Speech Translation task. Specifically, please refer to the Allowed Training Data and the Past Editions Development Data sections.

Text-to-Text Track (English to Japanese)

For training, you may use the parallel data and monolingual data available for the English-Japanese WMT20 news task. For development, you may use the IWSLT 2017 development sets and IWSLT 2021 development set.

English-to-Japanese simultaneous interpretation transcripts

You may also use simultaneous interpretation transcripts for the IWSLT 2021 development set for the comparison with human interpretation, under the terms of use written in README_before_download_enjaDevSI.txt.

Baseline Implementation and Example

English-to-German Speech-to-Text Translation

You can find a baseline and instructions on how to reproduce it here. Our final evaluation will be run inside Docker. To run an evaluation with Docker, first build a Docker image from the Dockerfile. Here is an example Dockerfile for the baseline:

FROM ubuntu:20.04


RUN apt-get update && apt-get install -y build-essential git python3 python3-pip libsndfile1
RUN git clone /fairseq
RUN pip3 install torch torchaudio vizseq soundfile sentencepiece sacrebleu=="1.5.1"
WORKDIR /fairseq
RUN pip3 install -e .
RUN git clone /SimulEval
WORKDIR /SimulEval
RUN pip3 install -e .
RUN ln -s /usr/bin/python3 /usr/bin/python

ENTRYPOINT simuleval \
    --agent $AGENT \
    --source $SRC_FILE \
    --target $TGT_FILE \
    --output $OUTPUT \
    --scores $EXTRA_AGENT_ARGS

Assuming your current directory contains the Dockerfile above, you can run the following commands to run the baseline evaluation inside Docker:

WORKDIR=<WORKING DIRECTORY> # `input` contains source/target files and wav files, `models` contains databin and the checkpoint
EXTRA_AGENT_ARGS="--data-bin $DATA --config $CONFIG_YAML --model-path $MODEL"

docker build -t iwslt2021_simulst_baseline:latest .
docker run -e AGENT="$AGENT" -e SRC_FILE="$SRC_FILE" -e TGT_FILE="$TGT_FILE" -e EXTRA_AGENT_ARGS="$EXTRA_AGENT_ARGS" -e OUTPUT="$OUTPUT" -v $FAIRSEQ:/SimulEval/fairseq -v $WORKDIR/input:/SimulEval/input -v $WORKDIR/models:/SimulEval/models -it iwslt2021_simulst_baseline

If you encounter a bus error similar to this issue, you can try adding --shm-size 8G to the docker run command.

When submitting your system, please make sure it works for the MuST-C dev and test sets. During the official evaluation, we will run the submitted system with the blind set.

English-to-Japanese Text-to-Text Translation

You can find instructions to train and evaluate an English-to-Japanese baseline system here.

System Submission

Participants are required to run the evaluation on the English-German dev and tst-COMMON MuST-C sets for the English-German track and on the IWSLT21 dev set for the English-Japanese track and report the results as part of the submission. This is to make sure that the submitted systems work so that organizers can run them as well. The submission files should be packed into a zip or tar.gz file and uploaded to dropbox prior to the deadline (Apr 23 2021 anywhere on earth). The submission files should include instructions on how to run the system in a README or file as well as all the necessary files (Docker image, checkpoints, vocabulary, etc.) for the organizers to be able to run the system.


Results are reported in the findings paper. We’ve also made available the system logs.




  • Katsuhito Sudoh (NAIST)
  • Satoshi Nakamura (NAIST)
  • Ondřej Bojar (Charles University)
  • Xutai Ma (Johns Hopkins University, Facebook)
  • Maha Elbayad (Facebook)
  • Changhan Wang (Facebook)
  • Juan Pino (Facebook)