Turn_taking_prediction_SWBD

Usage

import librosa
from espnet2.bin.asr_inference import Speech2Text

speech2text = Speech2Text.from_pretrained(model_tag="espnet/Turn_taking_prediction_SWBD")
# librosa resamples and mixes to one channel, so any file works; 16000 is
# what nearly every espnet recogniser is trained on - check this model's
# config if its audio is not 16 kHz
speech, rate = librosa.load("audio.wav", sr=16000, mono=True)
text, *_ = speech2text(speech)[0]
print(text)

ESPnet2 Turn taking model

espnet/Turn_taking_prediction_SWBD

This model was trained by “siddhu001” using swbd recipe in espnet.

Demo: How to use in ESPnet2

Follow the ESPnet installation instructions if you haven't done that already.

cd espnet
git checkout cea64abdeea5fa4f3da1a898be396e8c95c6e3ae
pip install -e .
cd egs2/swbd/asr1
./run.sh --skip_data_prep false --skip_train true --download_model espnet/Turn_taking_prediction_SWBD

Use the following Python code to run inference and obtain the probability of a turn-taking event every 40 milliseconds.

import soundfile
import os
import sys
from espnet2.bin.asr_inference import Speech2Text
speech2text = Speech2Text("exp/asr_train_asr_whisper_turn_taking_raw_en_word/config.yaml", "exp/asr_train_asr_whisper_turn_taking_raw_en_word/valid.loss.ave.pth",device="cuda", run_chunk=True)
audio, rate = soundfile.read(key)
print(speech2text(audio)[0][0])

RESULTS

asr_train_asr_whisper_turn_taking_target_raw_en_word

ROC_AUC

ASR config

<details><summary>expand</summary>

config: conf/train_asr_whisper_3_uselast.yaml
print_config: false
log_level: INFO
drop_last_iter: false
dry_run: false
iterator_type: sequence
valid_iterator_type: null
output_dir: exp/asr_train_asr_whisper_3_uselast_raw_en_word
ngpu: 1
seed: 0
num_workers: 1
num_att_plot: 3
dist_backend: nccl
dist_init_method: env://
dist_world_size: 8
dist_rank: 0
local_rank: 0
dist_master_addr: localhost
dist_master_port: 33429
dist_launcher: null
multiprocessing_distributed: true
unused_parameters: true
sharded_ddp: false
cudnn_enabled: true
cudnn_benchmark: false
cudnn_deterministic: true
collect_stats: false
write_collected_feats: false
max_epoch: 32
patience: null
val_scheduler_criterion:
- valid
- loss
early_stopping_criterion:
- valid
- loss
- min
best_model_criterion:
-   - valid
    - loss
    - min
keep_nbest_models: 10
nbest_averaging_interval: 0
grad_clip: 5.0
grad_clip_type: 2.0
grad_noise: false
accum_grad: 1
no_forward_run: false
resume: true
train_dtype: float32
use_amp: false
log_interval: null
use_matplotlib: true
use_tensorboard: true
create_graph_in_tensorboard: false
use_wandb: false
wandb_project: null
wandb_id: null
wandb_entity: null
wandb_name: null
wandb_model_log_interval: -1
detect_anomaly: false
use_adapter: false
adapter: lora
save_strategy: all
adapter_conf: {}
pretrain_path: null
init_param: []
ignore_init_mismatch: false
freeze_param:
- encoder
num_iters_per_epoch: 750
batch_size: 4000
valid_batch_size: null
batch_bins: 1000000
valid_batch_bins: null
train_shape_file:
- exp/asr_stats_raw_en_word/train/speech_shape
- exp/asr_stats_raw_en_word/train/text_shape.word
valid_shape_file:
- exp/asr_stats_raw_en_word/valid/speech_shape
- exp/asr_stats_raw_en_word/valid/text_shape.word
batch_type: folded
valid_batch_type: null
fold_length:
- 80000
- 150
sort_in_batch: descending
shuffle_within_batch: false
sort_batch: descending
multiple_iterator: false
chunk_length: 500
chunk_shift_ratio: 0.5
num_cache_chunks: 1024
chunk_excluded_key_prefixes: []
chunk_default_fs: null
train_data_path_and_name_and_type:
-   - dump/raw/train/wav.scp
    - speech
    - kaldi_ark
-   - dump/raw/train/text
    - text
    - text
valid_data_path_and_name_and_type:
-   - dump/raw/valid/wav.scp
    - speech
    - kaldi_ark
-   - dump/raw/valid/text
    - text
    - text
allow_variable_data_keys: false
max_cache_size: 0.0
max_cache_fd: 32
allow_multi_rates: false
valid_max_cache_size: null
exclude_weight_decay: false
exclude_weight_decay_conf: {}
optim: adam
optim_conf:
    lr: 0.0005
scheduler: warmuplr
scheduler_conf:
    warmup_steps: 500
token_list:
- <blank>
- <unk>
- C
- NA
- I
- BC
- T
- <sos/eos>
init: null
input_size: 1
ctc_conf:
    dropout_rate: 0.0
    ctc_type: builtin
    reduce: true
    ignore_nan_grad: null
    zero_infinity: true
    brctc_risk_strategy: exp
    brctc_group_strategy: end
    brctc_risk_factor: 0.0
joint_net_conf: null
use_preprocessor: true
use_lang_prompt: false
use_nlp_prompt: false
token_type: word
bpemodel: null
non_linguistic_symbols: null
cleaner: null
g2p: null
speech_volume_normalize: null
rir_scp: null
rir_apply_prob: 1.0
noise_scp: null
noise_apply_prob: 1.0
noise_db_range: '13_15'
short_noise_thres: 0.5
aux_ctc_tasks: []
frontend: null
frontend_conf: {}
specaug: null
specaug_conf: {}
normalize: null
normalize_conf: {}
model: espnet
model_conf:
    ctc_weight: 0.0
    lsm_weight: 0.1
    length_normalized_loss: false
    superb_setup: true
    num_class: 5
    ssl_input_size: 1024
    extract_feats_in_collect_stats: false
    use_only_last_correct: true
preencoder: null
preencoder_conf: {}
encoder: whisper
encoder_conf:
    whisper_model: medium
    dropout_rate: 0.0
    use_specaug: false
    specaug_conf:
        apply_time_warp: true
        time_warp_window: 5
        time_warp_mode: bicubic
        apply_freq_mask: true
        freq_mask_width_range:
        - 0
        - 40
        num_freq_mask: 2
        apply_time_mask: true
        time_mask_width_ratio_range:
        - 0.0
        - 0.12
        num_time_mask: 5
postencoder: null
postencoder_conf: {}
decoder: null
decoder_conf: {}
preprocessor: default
preprocessor_conf: {}
required:
- output_dir
- token_list
version: '202402'
distributed: true

</details>

Citing ESPnet

@inproceedings{
arora2025talking,
title={Talking Turns: Benchmarking Audio Foundation Models on Turn-Taking Dynamics},
author={Siddhant Arora and Zhiyun Lu and Chung-Cheng Chiu and Ruoming Pang and Shinji Watanabe},
booktitle={The Thirteenth International Conference on Learning Representations},
year={2025},
url={https://openreview.net/forum?id=2e4ECh0ikn}
}

@inproceedings{watanabe2018espnet,
  author={Shinji Watanabe and Takaaki Hori and Shigeki Karita and Tomoki Hayashi and Jiro Nishitoba and Yuya Unno and Nelson Yalta and Jahn Heymann and Matthew Wiesner and Nanxin Chen and Adithya Renduchintala and Tsubasa Ochiai},
  title={{ESPnet}: End-to-End Speech Processing Toolkit},
  year={2018},
  booktitle={Proceedings of Interspeech},
  pages={2207--2211},
  doi={10.21437/Interspeech.2018-1456},
  url={http://dx.doi.org/10.21437/Interspeech.2018-1456}
}