---
license: apache-2.0
task_categories:
- automatic-speech-recognition
language:
- en
size_categories:
- n<1K
tags:
- synthetic
- kokoro
- conversational
---

# Kokoro dialogue ASR set

500 single-speaker English clips of ~25-30 s, synthesized with
[Kokoro-82M](https://huggingface.co/hexgrad/Kokoro-82M) (voice `af_heart`) reading
procedurally generated spoken-monologue passages. Intended as **augmentation** for ASR
fine-tuning, not as a standalone training set.

| split | clips | hours |
|---|---|---|
| train | 406 | 3.14 |
| test | 94 | 0.72 |

## Fields

- `audio` — 16 kHz mono
- `transcription` — orthographic transcript, exactly the text that was synthesized
- `topic` — which passage template produced it (the split is held out by this)
- `voice`, `speed`, `duration_s`

## How it was made

Text comes from a passage generator: 12 everyday monologue
topics (a home repair, a commute, a spoken work update, a voicemail, ...), each an ordered
list of stages with interchangeable lines, plus slot banks for times, dates, money,
reference numbers and proper nouns — the tokens ASR models most often get wrong. Clip length is
closed-loop: word count is sized from a running words-per-second estimate and regenerated
until the audio lands in the 25-30 s window, so nothing is padded or truncated.

## Limitations

- **Synthetic speech.** Clean, single-mic, studio-style. A model trained only on this gets
  good at Kokoro, not at real acoustics — no reverb, overlap, codec artifacts, or genuine
  disfluency. Mix with real audio and augment.
- **One speaker.** Every clip is the same synthetic voice, `af_heart`. Fine for ASR,
  useless for diarization or speaker-change detection.
- **Templated text.** Sentence frames repeat across clips. The train/test split holds out
  whole topics to keep the eval honest, but vocabulary coverage is still narrow.