Aratako/Magpie-Speech-Orpheus-125k

Magpie-Speech-Orpheus-125k

A \~125k-sample synthetic speech dataset generated by applying the [Magpie](https://github.com/magpie-align/magpie) instruction-synthesis approach to the [Orpheus-TTS](https://huggingface.co/canopylabs/orpheus-3b-0.1-pretrained) LLM-based text-to-speech model, then decoding audio tokens with the [SNAC 24 kHz](https://huggingface.co/hubertsiuzdak/snac_24khz) codec.

• —Blog (EN): https://huggingface.co/blog/Aratako/magpie-speech
• —Blog (JA): https://zenn.dev/aratako_lm/articles/87d8988d44ba4d

This dataset is entirely synthetic: text prompts and audio tokens were produced by Orpheus-TTS and decoded to waveforms via SNAC.

What’s inside

• —Total duration: ~306.6 hours
• —Audio: 24 kHz FLAC, exposed via the audio column (🤗 Datasets Audio feature).
• —Core fields (column names):

• —id — zero-padded string ID
• —text — synthesized sentence to be spoken
• —audio — decoded waveform (24 kHz, mono FLAC)
• —n_audio_tokens — number of discrete tokens prior to decoding
• —frames — token frames (7 tokens/frame in Orpheus)
• —sr — sampling rate (should be 24000)
• —duration_sec — audio duration in seconds
• —text_hash — normalized text hash (dedup helper)
• —num_chars — visible character count
• —cps — characters per second
• —ASR & quality fields (added by post-processing):

• —asr_text — Whisper transcription
• —ref_no_filler, hyp_no_filler — normalized text used for error rates
• —cer, wer — character/word error rates (0–1)
• —clip_ratio — proportion of samples with |x| ≥ 0.999
• —dc_offset — mean amplitude offset
• —dnsmos_ovrl — DNSMOS overall quality score

Quick start

from datasets import load_dataset

ds = load_dataset("Aratako/Magpie-Speech-Orpheus-125k", split="train")
print(len(ds))
print(ds.features)          # shows the Audio feature for `audio`
print(ds.column_names[:12]) # peek column order

How it was generated (summary)

1. 1.Text synthesis (Magpie-style): Prompt the LLM up to (but excluding) the user text segment and let it continue to produce the “instruction” (the sentence to be spoken).
2. 2.Audio-token generation: Feed the synthesized text back into Orpheus-TTS to produce discrete audio tokens.
3. 3.Decoding to waveforms: Decode tokens to 24 kHz audio using SNAC (hubertsiuzdak/snac_24khz).
4. 4.Filtering:

• —Exact-match deduplication by normalized text hash
• —Remove top/bottom 10% by CPS (characters/sec)
• —Transcribe with [openai/whisper-large-v3](https://huggingface.co/openai/whisper-large-v3) and keep samples with WER ≤ 0.15 and CER ≤ 0.05
• —Drop samples with excessive clipping or DC offset
• —Score with [DNSMOS](https://github.com/microsoft/DNS-Challenge/tree/master/DNSMOS) and remove bottom 15%

For full scripts and rationale, see the blog posts (EN/JA).

Notes & caveats

• —Synthetic data: Voices and text are from model distributions and may not reflect real speakers or natural prosody across all cases.
• —ASR-based filtering: WER/CER thresholds are helpful but not perfect; manual spot-checking is recommended for critical use.
• —Domain balance: The text domain reflects what the LLM produced under the Magpie-style prompting; check domain fit for your tasks.

License

This dataset is distributed under the Llama 3.2 License.

While the Orpheus-TTS model used for synthesis is released under Apache 2.0, it relies on Llama 3.2 3B weights. Accordingly, I believe the applicable terms are those of the Llama 3.2 License. Therefore, this dataset inherits and is released under the Llama 3.2 License.

How to cite

@misc{magpie_speech_orpheus_125k,
  title        = {Magpie-Speech-Orpheus-125k},
  author       = {Aratako},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/datasets/Aratako/Magpie-Speech-Orpheus-125k}},
}