DograhAI-hq/tts-verbalization-audio

TTS Verbalization Benchmark — audio

7,560 speech clips: 1,080 English sentences spoken by each of seven open text-to-speech models. Every clip carries the sentence that produced it, the reference reading, both recognizer transcripts and both verdicts.

This is the audio behind dograh-hq/experiments, which publishes the corpus, the code, the results and a free verification script. The benchmark itself runs from transcripts and needs none of this audio. You want these files if you are checking a verdict by ear, training or evaluating a recognizer on hard verbalization cases, or studying how different models read dates, currency and symbols.

What the benchmark measures

Whether a model says a non-standard token correctly. Not audio quality, not naturalness. If the text says 05/20/2023, does the audio say "May twentieth, twenty twenty-three", or "zero five slash twenty slash two zero two three"?

Headline results, scored with OpenAI whisper-1:

model	correct	95% CI
kokoro-82m	0.61	0.58–0.64
orpheus-3b	0.54	0.51–0.57
qwen3-tts	0.49	0.46–0.52
chatterbox-turbo	0.49	0.46–0.52
chatterbox	0.44	0.41–0.47
chatterbox-multilingual	0.25	0.23–0.28
parler-tts	0.03	0.02–0.04

The smallest model is the most accurate. No model can say a hashtag: 0 correct out of 279. The best scored date result is 0.23, which is a floor rather than a measurement, for the reason below.

Layout

<model>/<item_id>.wav
metadata.jsonl

metadata.jsonl has one row per clip:

field	meaning
file_name	path to the clip, relative to the repo root
item_id	corpus item, e.g. syn-en-date-001
model, model_revision	which system spoke it, pinned by version hash
category	one of 27, e.g. Date, Currency, Hashtag or Mention
written	the text given to the model
spoken_reference	the correct spoken form
target_span	the part that is actually scored
whisper_transcript, whisper_correct	OpenAI whisper-1, the primary scorer
nova3_transcript, nova3_correct	Deepgram nova-3, a cross-check only

*_correct is null where a clip was excluded rather than scored: 47 lost to recognizer rate limits, 22 screened as unusable. They are excluded, not guessed at.

from datasets import load_dataset
ds = load_dataset("DograhAI-hq/tts-verbalization-audio", split="train")

Playing a clip directly over HTTP also works, which is what the write-up does:

https://huggingface.co/datasets/DograhAI-hq/tts-verbalization-audio/resolve/main/kokoro-82m/syn-en-date-001.wav

Where the transcript cannot settle it, and the audio can

A recognizer sometimes writes back the original text instead of the words it heard. "February twenty sixth two thousand nine" and "two thousand nine dash zero two dash twenty six" can both come back as 2009-02-26. One is a correct reading and one is not, and no amount of text processing separates them.

The corpus marks which items this can happen to: writeback_scores is false for the 125 items (875 clips) where echoing the written form must not count as correct. Those items are scored strictly, so a correct reading that gets written back is recorded as a miss. That makes the published numbers conservative, never generous.

281 of 7,509 scored clips (3.7%) are misses of exactly this kind. They are concentrated in three categories:

category	unresolvable	of scored	share
Biological Classification	95	279	34%
Date	74	278	27%
Version Numbers	68	278	24%

The effect on the per-model ranking is nil. If every one of those clips were in fact a correct reading, the order does not change: 0.61 to 0.66, 0.54 to 0.60, 0.49 to 0.53, 0.49 to 0.53, 0.44 to 0.48, 0.25 to 0.28, 0.03 to 0.04.

The effect on the Date category is large. orpheus-3b scores 0.20 and could be as high as 0.68; chatterbox scores 0.23 and could be 0.55. So read the date figures as a floor.

Settling it takes an ear, which is why this audio is published. Filter metadata.jsonl to rows where whisper_correct is false and whisper_transcript matches written, and listen.

Why nova-3 is a cross-check and not a scorer

Deepgram nova-3 expands abbreviations on its own, turning "mph" into "miles per hour", which repairs a real verbalization failure into a pass. numerals=false stops digit formatting but not lexical expansion. The two recognizers rank all seven models identically and disagree on 832 of the 7,497 clips both scored, so trust the ranking and not any single verdict. Those 832 are listed in the GitHub repo.

Licence and provenance

The metadata, the corpus text and the selection are CC BY 4.0 by Dograh. The corpus is generated by gen_corpus.py from a seed and a spec file, so no third-party text is involved.

The audio is model output. Each clip was generated by one of seven models, all released under permissive licences that state no restriction on generated audio:

model	licence
kokoro-82m	Apache 2.0
chatterbox, chatterbox-turbo, chatterbox-multilingual	MIT
qwen3-tts	Apache 2.0
orpheus-3b	Apache 2.0
parler-tts	Apache 2.0

Three things worth knowing before you build on this:

• Chatterbox clips carry a neural watermark. Resemble AI's Perth watermarker runs on every Chatterbox output. It is imperceptible and imposes no licence restriction, but roughly 3,240 clips here are watermarked and you will find it if you analyse them.
• orpheus-3b is built on Llama-3.2-3B. The model card is Apache 2.0, but the base carries Meta's Community Licence. Relevant if you train on these outputs: that licence requires a model trained on Llama outputs to have a name beginning with "Llama".
• parler-tts was trained on LibriTTS_R and MLS English. We did not verify how those dataset licences flow through to generated audio. Lower risk than the other two notes, but stated because it is unchecked rather than cleared.

If you are a rights holder for any model here and want a subset removed, open an issue on the GitHub repo and we will take it down.

Known limitations

• Not a word error rate. Every sentence is chosen to be hard.
• Models as hosted, not bare checkpoints. Wrapper settings measurably change results; version hashes are in model_revision.
• English only.
• One voice per model, listed in the GitHub repo's Setup section. parler-tts needs a written voice description and was given a neutral, slow one, because the default ends "She speaks very fast" and produced truncated gibberish.
• The hashtag reference says the symbol ("hashtag green break"). Many people would read it aloud without. That category measures whether a model speaks the symbol, not whether it is understandable.

Citation

@misc{dograh2026ttsverbalization,
  title  = {TTS Verbalization Benchmark: can open text-to-speech read a date out loud?},
  author = {{Dograh}},
  year   = {2026},
  url    = {https://github.com/dograh-hq/experiments/tree/main/benchmarks/tts-verbalization}
}