---
pretty_name: Three-Speaker Audio Dataset with Timbral Speaker Embeddings
language:
- en
size_categories:
- 1K<n<10K
task_categories:
- audio-classification
- automatic-speech-recognition
tags:
- speaker-embeddings
- speaker-verification
- educational
- synthetic-speech
- anomaly-detection
dataset_info:
  features:
  - name: audio
    dtype:
      audio:
        sampling_rate: 16000
  - name: text
    dtype: string
  - name: speaker_id
    dtype: string
  - name: wavlm_embedding
    list: float64
  splits:
  - name: train
    num_bytes: 2860951126.0
    num_examples: 4500
  download_size: 2645884238
  dataset_size: 2860951126.0
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# Three-Speaker Audio Dataset with Timbral Speaker Embeddings

A teaching dataset maintained by **AI-Academy**. It pairs single-speaker English
utterances with precomputed timbral speaker embeddings, and is intended for
coursework and exercises rather than for benchmarking or production systems.

The dataset deliberately contains one injected inconsistency; locating it is one of
the intended exercises (see [The injected anomaly](#the-injected-anomaly)).

## Overview

| Property | Value |
| --- | --- |
| Examples | 4,500 |
| Split | `train` only |
| Speakers | 3 (`kore`, `puck`, `andrew`), 1,500 utterances each |
| Total audio | 12.77 hours |
| Utterance duration | mean 10.22 s, median 9.00 s, range 0.72–54.48 s |
| Transcript length | 129,917 words in total, mean 28.9 words per utterance |
| Language | English |
| Embedding | 128-dimensional, `float64`, L2-normalised |
| Files | 6 Parquet shards of 750 rows each |

## Dataset structure

Each record has four fields:

| Field | Type | Description |
| --- | --- | --- |
| `audio` | `Audio(sampling_rate=16000)` | Mono 16-bit PCM WAV stored inline. The declared sampling rate is the decoding target; the stored files retain their native rates (see caveats). |
| `text` | `string` | The orthographic content of the utterance. For the synthetic speakers this is the text supplied to the synthesis system; for the human speaker it is the corresponding transcript. |
| `speaker_id` | `string` | One of `kore`, `puck`, `andrew`. |
| `wavlm_embedding` | `list<float64>` | 128-dimensional timbral speaker embedding of the utterance. |

Records are ordered by speaker and each shard contains exactly one speaker, so index
ranges and shard boundaries coincide:

| Index range | Speaker | Shards |
| --- | --- | --- |
| 0–1,499 | `kore` | `train-00000`, `train-00001` |
| 1,500–2,999 | `puck` | `train-00002`, `train-00003` |
| 3,000–4,499 | `andrew` | `train-00004`, `train-00005` |

This layout is convenient for selective loading but has consequences for how the data
may be partitioned; see the caveats below.

## Speakers and provenance

| Speaker | Origin | Utterances | Duration | Mean length | Native rate | Register |
| --- | --- | --- | --- | --- | --- | --- |
| `andrew` | Human speaker, recorded | 1,500 | 3.78 h | 9.08 s | predominantly 48 kHz, a minority at 44.1 kHz | Conversational, largely customer-support dialogue |
| `kore` | Synthesised with a Google Gemini text-to-speech voice | 1,500 | 4.59 h | 11.02 s | 24 kHz | Read prose: news, encyclopaedic and expository text |
| `puck` | Synthesised with a Google Gemini text-to-speech voice | 1,500 | 4.39 h | 10.54 s | 24 kHz | Read prose: news, encyclopaedic and expository text |

The mixture of one natural and two synthetic voices is intentional: it allows the same
corpus to serve both speaker-modelling exercises and exercises that contrast recorded
speech with contemporary synthesis.

## Speaker embeddings

The `wavlm_embedding` field was produced with
[`Orange/Speaker-wavLM-tbr`](https://huggingface.co/Orange/Speaker-wavLM-tbr) (the W-TBR
variant), a WavLM-large derivative that encodes the *timbral* traits of a voice while
suppressing prosodic ones. The procedure was:

1. decode the audio to 16 kHz, averaging channels to mono;
2. truncate the signal to the first 320,000 samples (20 seconds);
3. apply mean-variance normalisation and the model's statistics-pooling head.

The output is a unit-norm 128-dimensional vector, so the cosine similarity between two
embeddings reduces to their dot product. The authors of the model report an equal error
rate of 1.68 % on the VoxCeleb1-clean verification task at a decision threshold of
**0.472**, which gives a principled reference point for exercises that require one.

Two consequences are worth stating explicitly. First, because of step 2, the embedding
of any utterance longer than 20 seconds describes only its first 20 seconds. Second,
the pipeline is reproducible: recomputing embeddings from the stored audio recovers the
released vectors to a cosine similarity of approximately 0.998 or above, which makes
reproduction itself a usable laboratory exercise.

## Intended educational use

The dataset was assembled to support, among others, the following exercises.

**Vector operations and similarity.** The embeddings are unit-norm, low-dimensional and
cleanly separated by speaker, which makes them a suitable substrate for a first
treatment of cosine similarity and cosine distance, Euclidean distance on the unit
sphere and its relation to the cosine, class centroids and centroid-based assignment,
and the effect of normalisation on each of these.

**Speaker verification.** Constructing same-speaker and different-speaker trial pairs
from the released vectors permits a full verification pipeline: score distributions,
receiver operating characteristic and detection error tradeoff curves, equal error rate,
and the calibration of a decision threshold, which may then be compared against the
published value of 0.472.

**Classification and representation analysis.** Speaker identity is close to linearly
separable in embedding space, so the corpus supports both simple classifiers and the
diagnostic tooling around them: confusion matrices, learning curves, and projection by
principal component analysis, t-SNE or UMAP.

**Anomaly and label-noise detection.** See the following section.

**Signal processing.** The stored files retain heterogeneous native sampling rates, so
resampling, mono conversion, duration statistics, silence trimming and the extraction of
spectral features are all exercisable on the raw audio rather than on prepared arrays.

**Natural versus synthetic speech.** One speaker is a recorded human and two are
synthesised, which supports introductory work on synthetic-speech detection and on the
acoustic differences between the two sources.

**Speech recognition.** The paired `text` field permits evaluation or fine-tuning of
recognition models, with the caveat on transcript fidelity noted below.

**Data partitioning.** The corpus ships as a single `train` split. Designing a defensible
partition — and recognising why the naive sequential one fails here — is itself part of
the intended work.

## The injected anomaly

One utterance in the corpus has been deliberately corrupted. Its audio was replaced with
a recording of a different speaker who appears nowhere else in the dataset, while its
`speaker_id`, its `text` and its position in the corpus were left untouched. The
`wavlm_embedding` of that record was recomputed from the substituted audio, so the
embedding is consistent with the audio it accompanies but not with the label attached to
it.

Neither the affected speaker nor the affected index is disclosed here. Locating the
record is a well-posed exercise in outlier detection: the corrupted item is detectable in
embedding space, for instance by its cosine similarity to the centroid of its own
labelled speaker, and more than one detection strategy will succeed. The exercise
generalises to the practical problem of finding mislabelled records in a corpus that is
too large to audit by listening.

Instructors who require the ground-truth index should obtain it from the dataset
maintainers rather than from this document.

## Known characteristics and caveats

- **Speaker and domain are confounded.** The human speaker's material is conversational
  customer-support dialogue, whereas the synthetic speakers read expository prose. A
  classifier trained on this corpus may therefore separate the speakers partly on
  linguistic and channel grounds rather than on voice alone. Any claim about speaker
  modelling drawn from this data should be qualified accordingly.
- **Heterogeneous native sampling rates.** The declared feature rate of 16 kHz is a
  decoding target applied by the `datasets` library. The underlying WAV files are stored
  at their original rates, which differ between speakers and, within one speaker, between
  subsets.
- **Sequential ordering.** Because records are grouped by speaker, a partition taken by
  contiguous index yields splits that contain only one or two speakers. Shuffle, or
  stratify by `speaker_id`, before partitioning.
- **Wide duration range.** Utterances span 0.72 to 54.48 seconds. Batching without length
  bucketing or truncation will be inefficient, and embeddings of the longest utterances
  are subject to the 20-second truncation described above.
- **Transcript fidelity is not certified.** The `text` field has not been verified against
  the audio by forced alignment, and for the human speaker it should be treated as an
  approximate transcript rather than as a gold reference.
- **Balance.** The three speakers are balanced by utterance count (1,500 each) but not by
  total duration.
- **No held-out split.** Only `train` is provided.

## Loading

```python
from datasets import load_dataset

ds = load_dataset("aiacademy-kg/audio-3-speaker-datase-v2", split="train")

row = ds[0]
samples = row["audio"].get_all_samples()      # torchcodec API, datasets >= 4.0
waveform, sr = samples.data, samples.sample_rate
# on datasets < 4.0: waveform, sr = row["audio"]["array"], row["audio"]["sampling_rate"]

print(row["speaker_id"], row["text"][:60], len(row["wavlm_embedding"]))
```

Because the embeddings are unit-norm, similarity work needs no audio decoding at all:

```python
import numpy as np

emb = np.array(ds.select_columns("wavlm_embedding")["wavlm_embedding"])
spk = np.array(ds["speaker_id"])

similarity = emb @ emb.T                       # cosine similarity, embeddings are L2-normalised
centroids = {s: emb[spk == s].mean(0) for s in np.unique(spk)}
```

To avoid decoding the audio when only the metadata is required, load the Parquet files
directly with `pyarrow` and select the columns of interest.

## Attribution

The embedding extractor is the work of Orange SA and is distributed under
CC BY-SA 3.0. If the embeddings are used in published work, please cite:

> Gengembre, N., Le Blouch, O., Gendrot, C. (2024). Disentangling prosody and timbre
> embeddings via voice conversion. *Proc. Interspeech 2024*, 2765–2769.
> doi:10.21437/Interspeech.2024-207

```bibtex
@inproceedings{gengembre24_interspeech,
  title     = {Disentangling prosody and timbre embeddings via voice conversion},
  author    = {Nicolas Gengembre and Olivier {Le Blouch} and C\'edric Gendrot},
  year      = {2024},
  booktitle = {Interspeech 2024},
  pages     = {2765--2769},
  doi       = {10.21437/Interspeech.2024-207},
  issn      = {2958-1796},
}
```

## Licensing and intended scope

This corpus is released for teaching purposes. The synthetic portions were generated with
a third-party text-to-speech system and remain subject to that system's terms of use; the
recorded portion was contributed for educational use. Redistribution terms should be
confirmed with the maintainers before the data is used outside a teaching context, and
the corpus is not intended for the enrolment of speakers in biometric systems.