nytopop/expresso-improv

Info about this dataset

This is the improvised conversational subset of Expresso, machine segmented and transcribed with turn metadata preserved.

Models used:

• Silero VAD: initial segmentation of audio files
• Parakeet TDT 0.6B V2: transcription of each segment

An 80Hz highpass filter was applied to the source audio prior to segmentation to reduce incidence of the recording environment's low frequency noise.

In total, 27.4 hours of audio remains, as 16790 conversational turns. Some transcription errors are to be expected.

The Expresso Dataset

[paper] [demo samples] [Original repository]

Introduction

The Expresso dataset is a high-quality (48kHz) expressive speech dataset that includes both expressively rendered read speech (8 styles, in mono wav format) and improvised dialogues (26 styles, in stereo wav format). The dataset includes 4 speakers (2 males, 2 females), and totals 40 hours (11h read, 30h improvised). The transcriptions of the read speech are also provided.

You can listen to samples from the Expresso Dataset at this website.

Conversational Speech

The conversational config contains all improvised conversational speech as segmented 48KHz mono WAV.

Data Statistics

Here are the statistics of Expresso’s expressive styles:

---

Style	Read (min)	Improvised (min)	total (hrs)
angry	-	82	1.4
animal	-	27	0.4
animal_directed	-	32	0.5
awe	-	92	1.5
bored	-	92	1.5
calm	-	93	1.6
child	-	28	0.4
child_directed	-	38	0.6
confused	94	66	2.7
default	133	158	4.9
desire	-	92	1.5
disgusted	-	118	2.0
enunciated	116	62	3.0
fast	-	98	1.6
fearful	-	98	1.6
happy	74	92	2.8
laughing	94	103	3.3
narration	21	76	1.6
non_verbal	-	32	0.5
projected	-	94	1.6
sad	81	101	3.0
sarcastic	-	106	1.8
singing*	-	4	.07
sleepy	-	93	1.5
sympathetic	-	100	1.7
whisper	79	86	2.8
Total	11.5h	34.4h	45.9h

---

*singing is the only improvised style that is not in dialogue format.

Audio Quality

The audio was recorded in a professional recording studio with minimal background noise at 48kHz/24bit. The files for read speech and singing are in a mono wav format; and for the dialog section in stereo (one channel per actor), where the original flow of turn-taking is preserved.

License

The Expresso dataset is distributed under the CC BY-NC 4.0 license.

Reference

For more information, see the paper: EXPRESSO: A Benchmark and Analysis of Discrete Expressive Speech Resynthesis, Tu Anh Nguyen*, Wei-Ning Hsu*, Antony D'Avirro*, Bowen Shi*, Itai Gat, Maryam Fazel-Zarani, Tal Remez, Jade Copet, Gabriel Synnaeve, Michael Hassid, Felix Kreuk, Yossi Adi⁺, Emmanuel Dupoux⁺, INTERSPEECH 2023.