MithrilMan/teacher-in-the-loop

Teacher in the Loop: 8M-parameter students of Kokoro-82M

Try it in your browser · Audio samples · Code · Paper · PyPI

Three single-voice text-to-speech models of 8.07M parameters, distilled from Kokoro-82M with the revised recipe of the paper "Teacher in the Loop: Reliable Distillation of Kokoro-82M into 8M-Parameter Single-Voice Students, in English and Italian" (Fabio Angela, 2026). Each model speaks one of Kokoro's voices: af_heart (English), if_sara and im_nicola (Italian).

The architecture is Paradee's (code, model), unchanged: a text side of 4.23M parameters that predicts durations, pitch, energy and phoneme features, and a decoder of 3.85M parameters that turns them into a 24 kHz waveform. What changed is how the decoder is trained: the frozen teacher renders each training crop with the excitation the student receives ("teacher in the loop"), the decoder also learns the teacher's inverse-STFT head, and the final adversarial stage runs for 30,000 steps (the paper's "final recipe"). Each voice is one ONNX graph from phoneme ids to audio, 9.17 MB with int8 weights, running 24 to 26 times faster than real time on one CPU thread.

Code, training recipe and paper: https://github.com/MithrilMan/teacher-in-the-loop

Voices

UTMOS (utmos22_strong) on each voice's 200 held-out sentences. "Decoder alone": the decoder receives the teacher's own durations, pitch, energy and phoneme features. "End to end": text side and decoder, from phonemes, in PyTorch. "Int8": the exported graphs, ONNX Runtime on the CPU, end to end on the first 100 held-out sentences, int8 minus float32 (paired, mean ± standard error).

voice	language	Kokoro teacher	decoder alone (95% CI)	end to end	int8 − float32
af_heart	English	4.521	4.467 ± 0.006	4.474	−0.006 ± 0.001
if_sara	Italian	4.022	3.882 ± 0.025	3.895	−0.007 ± 0.011
im_nicola	Italian	3.867	3.696 ± 0.021	3.700	−0.010 ± 0.005

On the same 200 English sentences, Paradee's released decoder scores 4.371 (decoder alone). Read the Italian numbers with the caveats below: UTMOS was trained on English ratings.

Usage

With the titl package (Python 3.10 to 3.12):

pip install teacher-in-the-loop
titl --voice if_sara "Buongiorno a tutti." -o buongiorno.wav
titl --voice af_heart --speed 1.1 "Hello there." -o hello.wav   # --fp32 for the float32 graph

from titl import Voice, write_wav

voice = Voice.from_pretrained("af_heart")       # or "if_sara", "im_nicola"; int8 graph by default
audio = voice("Hello there. This voice has eight million parameters.", speed=1.0)   # float32, 24 kHz
write_wav("hello.wav", audio)

voice32 = Voice.from_pretrained("im_nicola", quantized=False)   # float32 graph

Voice.from_pretrained downloads <voice>/model_int8.onnx (or model.onnx) from this repository. Text is converted to phonemes as Kokoro does: Misaki 0.9.4 for English, Misaki with eSpeak NG for Italian, with the dental stops t̪ and d̪ mapped to plain t and d. Long text is read one sentence at a time, in chunks of at most 510 phonemes. English also needs spaCy's en_core_web_sm model, which Misaki downloads with pip on first use; to work offline or without pip, install it beforehand (python -m spacy download en_core_web_sm).

The graphs can also be used without the package, from phonemes:

import json
import numpy as np
import onnxruntime as ort
from huggingface_hub import hf_hub_download

repo = "MithrilMan/teacher-in-the-loop"
vocab = json.load(open(hf_hub_download(repo, "vocab.json"), encoding="utf-8"))
session = ort.InferenceSession(hf_hub_download(repo, "if_sara/model_int8.onnx"), providers=["CPUExecutionProvider"])

from titl.phonemes import Phonemizer           # or any source of Kokoro phonemes
phonemes = Phonemizer("it")("Buongiorno a tutti.")
ids = [0] + [vocab[c] for c in phonemes if c in vocab] + [0]
waveform = session.run(["waveform"], {"input_ids": np.array([ids], dtype=np.int64),
                                      "speed": np.array([1.0], dtype=np.float32)})[0][0]

ONNX contract

	name	type	shape	content
input	input_ids	int64	[1, T]	phoneme ids from vocab.json with the padding id 0 at both ends; at most 510 phonemes (T ≤ 512)
input	speed	float32	[1]	1.0 is the teacher's pace; above 1 speaks faster
output	waveform	float32	[1, N]	mono audio at 24 kHz

The excitation is random, so two runs on the same input differ slightly. The graphs contain no post-filter (see below).

Files

file	size	content
vocab.json	1 KB	Kokoro-82M's phoneme vocabulary (from its config.json), shared by all voices
<voice>/model_int8.onnx	9.17 MB	the voice as one graph, int8 weights with one scale per output channel (biases, norms and windows in float32), float32 arithmetic
<voice>/model.onnx	32.8 MB	the same graph with float32 weights
<voice>/text.safetensors	16.9 MB	PyTorch weights of the text side, training.student.TextStudent("s+mlp")
<voice>/decoder.safetensors	15.4 MB	PyTorch weights of the decoder, training.student_decoder.StudentDecoder("A")
<voice>/config.json		voice, language, sample rate, ONNX contract, phoneme settings, teacher revision, headline results
<voice>/export-validation.json		the exporter's CPU check of both graphs: output lengths, timing, int8 parity
LICENSE		the Apache License 2.0
SHA256SUMS		SHA-256 of every other file

The PyTorch files hold model weights only, no optimizer state. The ONNX graphs were exported from exactly these weights. To use them with the training and evaluation code of the repository, which reads checkpoints with a model key:

import torch
from safetensors.torch import load_file
from training.student import TextStudent
from training.student_decoder import StudentDecoder

text = TextStudent("s+mlp")
text.load_state_dict(load_file("if_sara/text.safetensors"))
decoder = StudentDecoder("A")
decoder.load_state_dict(load_file("if_sara/decoder.safetensors"))
torch.save({"model": decoder.state_dict()}, "decoder.pt")   # e.g. for python -m training.export

Training

• Teacher: Kokoro-82M v1.0, revision f3ff3571791e39611d31c381e3a41a3af07b4987, with the voice pack of the student's voice.
• Data: synthetic speech only. Kokoro-82M renders sentences from WikiText-103 for af_heart (11,999 sentences, the first 200 held out; the decoder used 5,799 training sentences) and 12,000 training and 200 held-out sentences from 1,203 articles of the Italian Wikipedia (snapshot 20231101.it of wikimedia/wikipedia, no article in both splits) for if_sara and im_nicola. No recorded human speech is used. The decoder trains on a pool of 150,000 random 64-frame crops rendered by the frozen teacher.
• Decoder, final recipe: stage one for 30,000 steps from scratch (spectral losses plus distillation of the teacher's inverse-STFT head, with the teacher's source module copied in and frozen), a warm restart of 20,000 steps, then the adversarial stages GAN10 (5,000 steps) and GAN3 (30,000 steps).
• Text side: Paradee's s+mlp student trained with the upstream text recipe (8,000 steps), not retrained for these models.
• Compute: one shared 32 GB GPU; about eleven to twelve hours of training per voice, plus about one hour and 62 GB of disk to render the pool.

python -m training.recipe VOICE in the code repository runs the whole recipe. The paper checked the training package on if_sara only (3.878 against 3.882 for the released decoder); the int8 cost varied between trainings there, so measure it for each new voice.

Evaluation and limitations

• One automatic metric. Almost every number is a UTMOS prediction. UTMOS was trained on English ratings; the Italian teachers themselves score only 4.02 and 3.87, and Kokoro's documentation grades both Italian voices C. Read the Italian numbers as a comparison between systems on the same voice, not as absolute quality. In English UTMOS is close to saturation: the released Paradee decoder, Paradee's paper model and all the paper's English decoders after the adversarial stages lie between 4.36 and 4.47.
• No human MOS. The listening evidence is two blind four-sentence comparisons in English by a single listener, the author, plus informal listening. A faint metallic ring added by the small decoder is the main audible defect known.
• Harmonic definition above 2 kHz. The students keep the teacher's harmonic contrast below 1 kHz, but between 2 and 4 kHz they reach 5.5 dB (af_heart) and 7.7 dB (if_sara) against the teacher's 10.5 and 11.8 dB: they fill part of the space between the upper harmonics with noise. UTMOS does not register this.
• Bounded by the teacher. The students imitate Kokoro, including its mispronunciations, its prosody and the errors of its text-to-phoneme conversion. Whisper large-v3-turbo transcribes them end to end with word error rates of 3.29%, 3.94% and 4.33% (teachers 3.32%, 4.09% and 4.45%), and their speaker similarity to the teacher's audio of the same sentence is 0.990, 0.995 and 0.993.
• Seeds and held-out set. Each voice was trained once, except if_sara, whose final recipe was repeated with two more seeds (3.882, 3.881 and 3.883). The 200 held-out sentences also served for the paper's design decisions, so they are a development set, not an untouched test set.
• Int8. The int8 graphs cost at most 0.01 UTMOS end to end, but they fail the exporter's parity gate (af_heart 2.45 vs 1.31 dB, if_sara 1.83 vs 1.37 dB, im_nicola 1.46 vs 1.04 dB of median log-spectral distance), because int8 rounding moves 1 to 4 of 546 token durations by one frame and the gate only compares cases where all durations match. The gate is reported in export-validation.json, not enforced. These are the graphs evaluated in the paper.
• No phase-locking filter. Unlike Paradee's released graph, which contains its 2–8 kHz phase-locking post-filter, these graphs have no post-filter: the filter lowers the score of decoders trained with this recipe (by 0.03 to 0.04 in English and 0.08 to 0.22 in Italian).
• Speed was measured on one desktop CPU (AMD Ryzen 9 9950X3D, ONNX Runtime 1.30) with Paradee's protocol.

License

The weights in this repository are released under the Apache-2.0 license. They are trained on the outputs of Kokoro-82M (Apache-2.0) with code derived from Paradee (Apache-2.0); vocab.json is Kokoro-82M's vocabulary. The graphs take phoneme ids and contain no third-party software. Converting text to phonemes with the titl package loads eSpeak NG and phonemizer-fork (both GPL-3.0-or-later) and num2words (LGPL-2.1), which pip installs as separate packages; anyone redistributing them together with the package must follow their licenses. The training sentences come from Wikipedia (CC BY-SA). Details: THIRD_PARTY_NOTICES.md in the code repository.

Citation

@misc{angela2026teacher,
  author       = {Fabio Angela},
  title        = {Teacher in the Loop: Reliable Distillation of {Kokoro-82M} into {8M}-Parameter Single-Voice
                  Students, in {English} and {Italian}},
  year         = {2026},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.23400040},
  url          = {https://doi.org/10.5281/zenodo.23400040}
}

@misc{mahendrakar2026paradee,
  author       = {Sahil Mahendrakar},
  title        = {{Paradee}: Distilling {Kokoro-82M} into an {8M}-Parameter Single-Voice Text-to-Speech Model},
  year         = {2026},
  url          = {https://github.com/sahilmahendrakar/paradee}
}