Thermostatic/neblinia-speech-preview-1.0

NeblinIA-Speech preview-1.0 (whisper-large-v3)

A foundational ASR model for Mexican Spanish + 23 Mexican Indigenous languages (Nahuatl, Mixtec, Zapotec, Mazatec, Zoque, Amuzgo, Chinantec, Tlapanec, Triqui, Totonac, Mixe, across 6+ language families), by the NeblinIA lab.

Results (MEXA contamination-resistant benchmark)

model	WER	CER
preview-1.0 (whisper-large-v3, this model)	54.4	24.2
preview-0.1 (whisper-large-v3-turbo + RL)	59.0	26.5

Scored on a private, held-out test set (5,925 clips) the model cannot have trained on, with the identical decoding protocol for every model (faster-whisper, beam 1, temperature fallback).

Read CER, not just WER. These languages have no standardized orthography, so references spell the same spoken word several ways (tone marks, vowel length, tz/ts/z, word segmentation). An error audit showed the model is acoustically strong (CER 24, about 76% of characters correct) while WER is inflated by orthographic-convention mismatch, not mishearing. WER overstates the real error.

What it is

• Base: openai/whisper-large-v3 (the full 32-layer decoder). The decoder capacity was the key lever: it beat the 4-layer turbo base (with RL) by about 5 WER points, before any RL.
• Training: LoRA SFT on a broad multilingual manifest (about 97k clips) built only from open-licensed data (Omnilingual ASR CC BY 4.0, Common Voice v26 CC0, CIEMPIESS CC BY-SA).
• Files: ct2/ (CTranslate2, for faster-whisper, the format these numbers come from) and lora/ (the LoRA adapter, applies on top of openai/whisper-large-v3).

Usage (faster-whisper, recommended)

from faster_whisper import WhisperModel
m = WhisperModel("Thermostatic/neblinia-speech-preview-1.0", revision="main")  # or local ct2/ dir
segs, _ = m.transcribe("audio.wav", beam_size=1)
print("".join(s.text for s in segs))

Reproducibility

• Code + benchmark + data recreation: https://github.com/Sekinal/neblinia-speech (see docs/RECREATE_DATA.md for the full open-source data pipeline) and https://github.com/Sekinal/mexa-benchmark (deterministic split + fingerprint registry).
• Honest log of everything tried, including the negative results (byte-level and ByT5 speech-LLM both loop; GSPO RL helps weak models but over-generates on this strong base; the open-data ceiling): docs/findings.md in the GitHub repo.

Limitations

• Zapotec and other data-dark languages stay weak: there is no open audio for them, so this is a data limit, not a model limit.
• WER near 20 is likely unreachable while the reference orthography itself is inconsistent. The honest content-accuracy metric is CER, and an orthographic-normalization protocol is in progress.
• Tone is phonemic in the Oto-Manguean languages; the model does not yet model it reliably.

License

CC BY 4.0 (matching the dominant training source, Omnilingual ASR). Some training sources are CC0 and CC BY-SA; see the GitHub speech-data-research repo for the full license audit.