Expresso — audio + text

A faithful re-publication of the official Expresso dataset (Nguyen et al., Interspeech 2023) as a loadable HuggingFace audio dataset, sourced directly from FAIR's official tar.

⚠️ License: CC-BY-NC-4.0 — non-commercial use only.

Configs

• read — 11.6k mono read-speech utterances with human transcripts.
• conversational — ~15.9k mono per-utterance turns derived from the stereo conversational dialogues, transcribed with Whisper Large V3 Turbo.

read config

11.6k mono utterances at 48 kHz / 24-bit, fully transcribed by humans.

	train	dev	test
rows	10,388	628	588

Schema

Column	Type	Notes
id	string	e.g. ex01_confused_00001; longform chunks: ex01_default_longform_00001__0-16.49
audio	Audio @ 48 kHz mono
text	string	human-written transcription (mixed case, with punctuation)
speaker_id	int32	1–4
style	string	one of: default, confused, enunciated, happy, laughing, narration, sad, whisper
substyle	string	finer-grained label, e.g. default_emphasis, default_essentials, default_longform, narration_longform
corpus	string	base (short utterances) or longform (multi-minute readings)
start_s	float32	null for full-file rows; chunk start for longform
end_s	float32	null for full-file rows; chunk end for longform

Splits

We follow the official Expresso train/dev/test splits, with one TTS-oriented deviation:

• base read (~11,600 utterances): full-file rows, no slicing — official splits applied as-is.
• longform read (8 source files: default_longform, narration_longform × 4 speakers): kept as full files in train only. The official Expresso splits slice each longform file into 3 non-overlapping chunks (60 s for dev/test, the rest for train) for resynthesis benchmarking. Those chunks don't align with the full-file transcripts, so for TTS/ASR we keep the longform audio + transcript intact and place the full files in train only. If you need the official chunked benchmark, see original_metadata/splits/.
• singing is intentionally excluded (only 12 wavs total, not in official splits).

All rows have aligned (audio, text) pairs.

Style coverage per speaker

All 4 speakers have all 8 styles, with these caveats:

• narration is longform-only for all speakers (1 file each).
• default includes the substyles default, default_emphasis, default_essentials, default_longform.

---

conversational config

~15.9k per-utterance mono turns derived from the official 339 stereo dialog files. Each row is one speaker's turn at a known time range within the source file, transcribed by Whisper.

	train	dev	test
rows	~14.8k	~520	~515
audio	~29 h	~50 min	~51 min

Schema

Column	Type	Notes
id	string	e.g. ex01-ex02_default_001__ch1_23.88-28.14
audio	Audio @ 48 kHz mono	the VAD-extracted turn from one channel
text	string	Whisper Large V3 Turbo transcript (mixed case + punctuation)
speaker_id	int32	this channel's speaker (1–4)
style	string	this channel's expressive style
other_speaker_id	int32	partner's speaker id
other_style	string	partner's expressive style
source_file_id	string	e.g. ex01-ex02_default_001 (the stereo source)
channel	int32	1 or 2
start_s	float32	turn start within source file (after VAD ∩ split clip)
end_s	float32	turn end

How it was built

1. Parse the official splits/{train,dev,test}.txt time-window assignments per source file.
2. Intersect each split window with VAD_segments.txt (per-channel pyannote turns) — turns straddling the dev/test boundary are clipped to the split window so dev/test never leak into train.
3. Slice the stereo source file → mono channel → 48 kHz mono turn.
4. Transcribe with openai/whisper-large-v3-turbo, with anti-hallucination decoding (no_repeat_ngram_size=4, repetition_penalty=1.2, condition_on_prev_tokens=False) and pre-resampled to 16 kHz.

Turn filtering

• Min duration: 0.3 s. Sub-300ms VAD turns (mostly backchannels and clicks) are dropped.
• Max duration: 28 s. Long turns are split into ≤28 s pieces (Whisper's context is 30 s).

Style coverage

26 styles total in the conversational subset, including styles not present in read: angry, animal, awe, bored, calm, desire, disgusted, fast, fearful, nonverbal, projected, sarcastic, sleepy, sympathetic, plus mixed pairs like animal-animaldir and child-childdir (where the two channels carry different styles — one row's style and other_style will differ).

ASR quality (validated against read ground truth)

We benchmarked Whisper Large V3 Turbo on 210 human-transcribed read utterances spanning all 7 transcribed read styles. Per-style WER:

confused	default	sad	happy	enunciated	laughing	whisper	overall
0.96%	1.67%	2.00%	2.76%	3.18%	4.98%	5.31%	3.00%

ASR errors are highest on whisper and laughing styles (the toughest acoustic conditions), but still under 6% WER. Conversational rows are expected to track the same per-style quality.

Caveats

• Transcripts are machine-generated — expect a small error rate, especially on whispered/laughing/animal-style turns.
• Mixed-style pairs (animal-animaldir, child-childdir, sad-sympathetic and reversals) — speakers in the two channels carry different styles. Ground-truth styles are encoded per-row in style (this channel) and other_style (partner).

---

Sidecar files

The original FAIR metadata is uploaded under original_metadata/:

• original_metadata/README.txt, LICENSE.txt — official Expresso documentation
• original_metadata/read_transcriptions.txt — per-file transcripts (tab-separated)
• original_metadata/VAD_segments.txt — per-channel VAD timings for the conversational subset (used to derive the conversational config)
• original_metadata/splits/{train,dev,test}.txt, splits/README — official split definitions

Quick start

from datasets import load_dataset

# Pick a config — there is no default
read = load_dataset("shangeth/expresso", "read",          split="train")
conv = load_dataset("shangeth/expresso", "conversational", split="train")

ex = read[0]
print(ex["id"], "|", ex["style"], "|", ex["text"])
print(ex["audio"]["array"].shape, "@", ex["audio"]["sampling_rate"], "Hz")

# Filter conv to mixed-style pairs (cross-style modeling)
mixed = conv.filter(lambda x: x["style"] != x["other_style"])
print(f"{len(mixed)} cross-style turns")

# Per-style coverage
from collections import Counter
print(Counter(conv["style"]).most_common(10))

Reproducing this dataset

# Download the official Expresso tar (~36 GB) and extract:
mkdir -p data && cd data
curl -L https://dl.fbaipublicfiles.com/textless_nlp/expresso/data/expresso.tar | tar -xf -
cd ..

# Build + push:
python expresso_audio.py        --repo_id shangeth/expresso --private  # read config
python expresso_conversational.py --repo_id shangeth/expresso --private  # conversational config

See github.com/shangeth/wren-datasets for the full extraction code.

Citation

@inproceedings{nguyen2023expresso,
  title     = {Expresso: A Benchmark and Analysis of Discrete Expressive Speech Resynthesis},
  author    = {Nguyen, Tu Anh and Hsu, Wei-Ning and D'Avirro, Antony and Shi, Bowen and
               Gat, Itai and Fazel-Zarani, Maryam and Remez, Tal and Copet, Jade and
               Synnaeve, Gabriel and Hassid, Michael and Kreuk, Felix and Adi, Yossi and Dupoux, Emmanuel},
  booktitle = {Interspeech},
  year      = {2023}
}

@misc{wren2026,
  title  = {Wren: A Family of Small Open-Weight Models for Unified Speech-Text Modelling},
  author = {Shangeth Rajaa},
  year   = {2026},
  url    = {https://github.com/shangeth/wren}
}

License

CC-BY-NC-4.0 — non-commercial use only. See original_metadata/LICENSE.txt.