Fluent Books: small Spanish voices (ONNX)

Ten small text-to-speech models for the Fluent Books reader app (read a Spanish book aloud, on the phone's processor). Each file is one voice: the voice is baked into the weights, the input is a sequence of Kokoro phoneme ids and a speed, the output is a 24 kHz waveform and the duration of every token (waveform [1,S], durations [1,T], S = 600 * sum(durations)).

They are students distilled from the Spanish voices of Kokoro-82M (ef_dora, em_alex, em_santa): one student per voice, trained on a mix of Castilian and Latin American phonemes (the dialect is chosen by the phonemes you give: theta and ll for Castilian, seseo and yeismo for Latin American). Two recipes: A (8.6 MB) and L60 (12.5 MB, 60 harmonics, males only); the name ends with the weight of the phase critic in the last training stage (ph3, ph10). The metrics could not tell the two weights apart: pick by ear.

How to use

Phonemes: the rules in es_g2p_rules.py of the Fluent Books repository (a Swift port is in the app, SpanishPhonemizer) imitate espeak-ng's es output, which is what Kokoro was trained on; the symbols must be in Kokoro's vocabulary (vocab.json, 114 symbols). input_ids = [0] + ids + [0] (int64, shape [1,N]), speed = [1.0] (float32, shape [1]).

Measured

Measured on 60 held-out sentences (vn_eval_x.sh, Whisper large-v3 es; the teacher's own WER is in the DONE files) and by the detector (tools/distill/artifact_card31.py, the reference = the teacher recordings of the same voice). The detector's thresholds are English-voice thresholds; em_santa has a weak teacher (UTMOS 2.89, UTMOS is trained on English), so its sigmas are small for the wrong reason. The detector measures, the ear decides.

Measured on 60 held-out sentences (Whisper large-v3 es; the teacher's own WER 1.8-2.4 %); every model has 0 contract violations. A second check before publication (20 sentences per model, 10 per dialect, phonemes of the Swift port through onnxruntime) gave 3.0-3.9 % WER, of which 3.0 % are three reference artifacts (Vd., mas, a name), and one Whisper hallucination at the end of a clip.

Open points (read before you rely on these)

• Data of the teacher. The students were trained on speech synthesized by the Kokoro-82M voices ef_dora, em_alex, em_santa, and on sentences from public-domain books (Project Gutenberg, authors who died before 1940). Kokoro-82M is Apache-2.0; where the data of its Spanish voices comes from is not confirmed by the authors of this repository. The weights are stated as Apache-2.0 on the reading of those sources.
• No native listener yet. The dialect (theta, seseo, yeismo) and the naturalness of the stress are not measured by any automatic metric; a native speaker has not listened to these models.
• The detector's thresholds (31 characteristics of "metal/noise") come from English voices and are only a guide for Spanish; UTMOS is trained on English and under-rates Spanish.
• Text must be normalised first (numbers, Roman numerals, dates); the rules read letters and the punctuation , . ; : ! ? - ... " " only.

Trained 7 November 2026. Source: the Fluent Books project (tools/distill, docs/spanish-speech-engine-research-2026-11-05.md).