日本語要約: tiny-voca-ja は、英語 TTS モデル tiny-voca(n4ze3m)を日本語で ゼロから学習し直した派生モデルです。パラメータ 4.5M、声は F2(女性)と M2(男性)の 2 声、出力は 24 kHz モノラル。学習データは Qwen3-TTS で生成した合成音声 16,321 クリップ(16.6 時間)で、実在の話者の録音は含まれません。CPU で実時間より速く 動きます。未学習16文の正式評価では CER 平均 6.6%(数字・単位の表記を読みにそろえた正規化後は 1.6%)でした。日付・数字の読み誤りの主因はモデルではなく、読み処理(フロントエンド)側にありました。2026年11月20日に、読み処理を pyopenjtalk-plus にそろえて再注釈したデータで追加学習した v2(114,000ステップ)を公開しました。同じ16文評価で、正規化後CERは 1.6% から 1.0% に改善しています。2026年11月21日には、英語と日英混在の教師データを追加した v3(134,000ステップ、フロントエンド identity は oj-en-1)を公開しました。日本語16文評価は v2 と同値(正規化後CER 1.0%、UTMOS 3.70)を維持し、未学習の英語40文では WER 2.8%(実質的な単語誤りは40文中1語)でした。
tiny-voca-ja is a Japanese text-to-speech model with two voices, F2 (female) and M2 (male). It is a from-scratch retrain of tiny-voca — the same architecture and recipe by Muhammed Nazeem — on Japanese data. It is not a fine-tune of the English weights: no pretrained weights are part of the model. The complete model, including the waveform generator, has 4.5M parameters and was trained on about 16.6 hours of synthetic speech. It runs faster than real time on CPU. A fine-tuned v3 (step 134,000, 2026-11-21) is the current checkpoint in this repository; see Fine-tune v3 below. It keeps the v2 Japanese scores and additionally reads English and mixed Japanese–English text. The v1 and v2 checkpoints remain available in the revision history.
| Parameters | 4,538,827 |
| Voices | F2, M2 (Japanese) |
| Output | 24 kHz mono |
| Input | Japanese text via pyopenjtalk-plus; English and mixed Japanese–English via the merged oj-en-1 front end (see Fine-tune v3) |
| Pretrained weights in the model | None |
| License | Apache-2.0 |
The parameter count is 1,536 higher than the English model (4,537,291) because the phoneme inventory grew from 76 to 84 symbols (8 Japanese additions).
| File | Size | Contents |
|---|---|---|
tiny-voca-ja.pt |
18.2 MB | PyTorch weights, float32 (inference checkpoint, format 4; v3, step 134,000) |
onnx/model.onnx |
21.4 MB | ONNX export, float32, opset 17 |
config.json |
Model settings, phoneme vocabulary, voices, front-end identity | |
samples/ |
The audio samples on this page, synthesized by this model | |
metrics.jsonl |
11.0 MB | Full training log of the v1 run |
metrics-finetune-v2.jsonl |
1.1 MB | Training log of the v2 fine-tune run |
metrics-finetune-v3.jsonl |
2.1 MB | Training log of the v3 fine-tune run |
SHA256:
tiny-voca-ja.pt: 0bae8a003b7c7b6e0d7cb16bcffaf927815329e6495c4b3509d28483eaee7e67onnx/model.onnx: 0ecb7542360d78e90a1571e9784c7b6474fbfde8cf43d8bbed12d54bfdb280c7The v1 ONNX export was verified against the PyTorch model on a Japanese
sentence for both voices: identical durations, and a waveform difference of
at most 0.16% of the peak amplitude (the noise signal is drawn by each
runtime, as in the English model). For v2, the web-app ONNX export of the
same checkpoint (same weights, a different graph wrapper) was verified
against PyTorch synthesis with correlation 1.000000 and a maximum waveform
difference of 0.0005. For v3, the web-app ONNX export was verified the same
way (correlation 1.000000 on all four fixture entries, maximum waveform
difference 0.0021), and the stock ONNX graph in onnx/ was smoke-checked
against PyTorch: identical durations and output length.
Synthesized on a CPU-only Linux machine (no GPU). CER is the character error rate of a Whisper small (multilingual) transcription against the input text, computed in a kana-normalized space so that kanji spelling variants are not counted as errors. See Evaluation for the caveats — this is four sentences, not a benchmark.
| # | Voice | Text | CER |
|---|---|---|---|
| 1 | F2 | こんにちは、これは日本語の音声合成のテストです。 | 0.0% |
| 2 | F2 | はじめまして。私は日本語で話す、小さな音声合成モデルです。 | 0.0% |
| 3 | M2 | 今日は十一月十九日、木曜日です。東京の天気は、晴れのち曇りでしょう。 | 0.0% |
| 4 | F2 | 人工知能が、日本語の文章を読み上げます。声の大きさも、速さも、自由に変えられます。 | 0.0% |
Sample 3 was scored 29.2% in the original spot check; that figure later turned out to be mostly a scoring artifact, and the same audio scores 0.0% once numeral notation is normalized on both sides — see the correction in Evaluation.
v3 checkpoint note. The current checkpoint (v3, step 134,000) was
fine-tuned for the merged front end oj-en-1 (pyopenjtalk-plus for
Japanese, eSpeak NG en-us phonemes for English), so the stock CLI below
— which targets the v1 eSpeak front end — will refuse it as a front-end
mismatch. Use the OJ adapter in the repository
(frontend-swap/oj_frontend.py) for Japanese, or the web app
tiny-voca-ja-web;
the Japanese component of oj-en-1 is identical to v2's oj-1. The v1
checkpoint in the revision history works with the CLI as written.
The Japanese front end uses eSpeak NG with the ja voice, so the checkpoint
only speaks Japanese text. You need the espeak-ng system package and the
code from the Japanese recipe repository.
sudo apt install espeak-ng # or your platform's equivalent
pip install git+https://github.com/gachimoto-code/tiny-voca-ja
hf download gachimoto/tiny-voca-ja --local-dir tiny-voca-ja
tiny-voca-speak \
--checkpoint tiny-voca-ja/tiny-voca-ja.pt \
--voice F2 \
--text "こんにちは、これは日本語の音声合成のテストです。" \
--output hello.wav
(python -m tiny_voca.infer with the same arguments works too.)
| Option | Default | Meaning |
|---|---|---|
--text-file |
Read the text from a file instead of --text. |
|
--voice |
F2 |
F2 or M2 |
--speed |
1.0 | Speaking rate |
--pitch-shift |
0 | Added to log F0. 0.058 is about one semitone. |
--pitch-range |
1.0 | Scales pitch movement around the utterance mean. |
Text of any length works. The command reads it one sentence at a time and
joins the audio. The checkpoint records the front end it was trained with;
inference refuses to run with a mismatched front end unless
--allow-frontend-mismatch is given.
Synthesis measured on a small shared CPU-only VM: 2.8–6.3 seconds per sentence including model load, for 4.4–8.1 seconds of audio — faster than real time, but this VM is far from a benchmark machine.
The graph takes phoneme ids and stress values and returns the waveform.
Text is converted to ids outside the graph with the Japanese front end
(eSpeak NG ja + pykakasi); the ids are the positions in vocab in
config.json. Inputs and outputs are the same as the English model:
inputs tokens, stress, voice (0 = F2, 1 = M2), speed, pitch_shift,
pitch_range; outputs audio (24 kHz float32) and durations. The model
reads one utterance per call, up to 512 tokens and 32 seconds. The audio can
peak a little above 1.0, so scale it down before saving 16-bit samples.
Text is converted to phonemes with a stress value for each. A transformer encoder (4 blocks) reads the phonemes, a duration predictor gives each one a length in frames, and two frame-level transformers (2 and 5 blocks) predict pitch, voicing and energy and then the parameters of the waveform generator. All blocks use 192 channels, 4 heads and rotary position embeddings.
The waveform is produced by a source-filter synthesizer instead of a neural vocoder. A band-limited pulse train at the predicted pitch and a noise signal are each shaped by a time-varying filter, whose cepstrum the network predicts once per frame (256 samples, 10.7 ms). This part has 25,862 parameters.
The Japanese model uses the same components as the English one; only the phoneme embedding grows with the 84-symbol inventory.
| Data | gachimoto/tiny-voca-ja: 16,321 clips, 16.6 hours, two voices; 11,561 clips (11.5 hours) passed annotation and were used for training (10,884 train / 332 validation / 345 test) |
| Steps | 104,000 |
| Hardware | One RTX 4090 (rented, RunPod), about 5.3 hours |
| Final loss | 10.717 (base loss 6.494) at step 104,000 |
| Losses | Multi-resolution spectral losses; adversarial loss with multi-period and multi-resolution discriminators after step 8,000 (gan-ratio 0.7) |
The training audio is synthetic. Both voices were created from text descriptions with Qwen3-TTS VoiceDesign, and the recordings were generated with Qwen3-TTS 1.7B Base. No recording of a real speaker is involved.
Phoneme durations come from forced alignment with
prj-beatrice/japanese-hubert-base-phoneme-ctc-v2, with labels from
pyopenjtalk-plus. This differs from the English recipe on purpose: the
eSpeak-based aligner used there (facebook/wav2vec2-lv-60-espeak-cv-ft)
accepted 0 of 150 Japanese smoke-test clips — it has not learned Japanese
acoustics — so a Japanese phoneme CTC model was used instead. 11,561 of the
16,321 clips (70.8%) passed the alignment checks.
The v1 checkpoint was fine-tuned for +10,000 steps (114,000 total) on the
same synthetic corpus, re-annotated end to end: pyopenjtalk-plus labels
with forced alignment by the beatrice phoneme recognizer, mapped to the
84-symbol inventory. 15,558 of the 16,321 clips passed the checks
(14,652 train / 453 validation / 453 test). The front-end identity of the
checkpoint therefore changed from eSpeak (v3) to oj-1
(pyopenjtalk-plus, g2p_prosody mode).
One honest caveat: because of a learning-rate schedule quirk (the decay term used the absolute step), the effective learning rate was about 2e-6 — roughly 1/40 of the intended 1e-4. The improvement below was obtained anyway, at about $2 of rented GPU time.
Samples from the v2 evaluation set (the 16-sentence formal evaluation, matched front end):
The v2 checkpoint was fine-tuned for another +20,000 steps (134,000
total) on the v2 data plus new teacher data for English and mixed
Japanese–English — the goal of this round. New sentences (172 additional
Japanese, 600 English, 300 mixed) were synthesized with the same
Qwen3-TTS teacher and screened with Whisper; annotation accepted 300
Japanese, 1,045 English and 494 mixed clips. English labels are eSpeak NG
(en-us) phonemes with forced alignment; mixed sentences were aligned by
a span-splitting aligner (Japanese spans with the beatrice aligner,
English spans with eSpeak CTC). The 84-symbol inventory already contained
the English phonemes, so no vocabulary change was needed; the merged
front-end identity is oj-en-1. Training rows: 17,536 (v2 train 14,652 +
JA 300 + EN 1,045 ×2 oversampled + mixed 494).
The v2 learning-rate quirk was compensated in the arguments (base LR argument 0.00538), giving an effective peak LR of ≈1e-4 as originally intended. Results on held-out evaluation sets (Whisper large-v3):
The parameter count is unchanged (4,538,827). This round cost $4.00 of rented GPU time in total (≈ ¥637 at ¥159.5/$).
A mel + HiFi-GAN vocoder (see Vocoder pilot in the GitHub README) can be applied as post-processing; measured on the v3 evaluation samples it improves the male voice M2 (UTMOS +0.50 on average over 8 clips) but degrades the female voice F2, so applying it to M2 only is recommended.
Character error rate of synthesized speech, compared in a kana-normalized space so that kanji spelling variants are not counted as errors. Two checks have been run, both synthesized locally on CPU: a 4-sentence sanity check transcribed with Whisper small (multilingual), and a formal evaluation of 16 unseen sentences transcribed with Whisper large-v3:
| Sentences | Clips | CER |
|---|---|---|
| New sentences, not in the dataset | 4 | 7.3% mean (0.0 / 0.0 / 29.2 / 0.0) with the original kana-only scoring; 0.0% mean after numeral normalization |
| Formal evaluation, 16 unseen sentences (Whisper large-v3) | 16 | 6.6% mean raw, 1.6% after numeral/unit normalization |
| Formal evaluation, 16 unseen sentences, v2 fine-tune with the matched pyopenjtalk-plus front end (Whisper large-v3) | 16 | 6.0% mean raw, 1.0% after numeral/unit normalization; UTMOS 3.73 (v1 with the same front end: 1.6% normalized, UTMOS 3.67) |
Formal evaluation, 16 unseen sentences, v3 fine-tune with the matched oj-en-1 front end (Whisper large-v3) |
16 | 6.0% mean raw, 1.0% after numeral/unit normalization; UTMOS 3.70 — identical CER to v2 |
English evaluation, 40 unseen sentences, v3 (Whisper large-v3, language="en") |
40 | WER 2.84% / CER 2.75%; UTMOS 4.05 |
| Mixed Japanese–English evaluation, 20 unseen sentences, v3 (Whisper large-v3) | 20 | Katakana-folded CER 10.88%; UTMOS 3.71 |
The 4-sentence row is a sanity check, not a benchmark. The formal evaluation has since been run (2026-11-19): 16 unseen sentences covering dates in both numeral styles, times, money, counters, place and personal names, a phone number, decimals, homophones and one long sentence, transcribed with Whisper large-v3. The stock front end scored 6.6% mean CER raw and 1.6% after normalizing numeral and unit notation on both sides; 13 of the 16 sentences scored 0.0% normalized. No listening test has been run.
Correction (2026-11-19). Sample 3's 29.2% was mostly a measurement artifact: the ASR writes dates as digits ("11月19日") while the reference is read as kanji numerals, and the kana scorer drops digits, so matching sounds were counted as errors on one side only. Normalized, that sentence scores 0.0% — the audio was correct. The genuine reading errors were in Arabic-numeral date sentences (「今日は」 read as the greeting 「こんにちは」, a wrong weekday); an experiment in the GitHub repository (experiments/frontend-swap) shows they come from the front end, not the model: with pyopenjtalk-plus readings the same checkpoint scores 0.0% on that sentence, and 1.6% mean normalized CER over the 16-sentence set.
oj-en-1 front end, but multi-word English phrases and proper nouns in
mixed text are still unreliable (see Fine-tune v3).About $20 of rented GPU time in total (estimated from pod hours × the $0.89/h rate; not reconciled against the RunPod billing screen). Roughly: $9.18 for dataset generation, $4.69 for the final training run, the rest for voice design, environment setup and the failed annotation attempts. The English original cost $9.56; the Japanese version cost about twice that, mainly because of the aligner replacement and the higher per-hour rate of the GPU that was actually available. The v3 fine-tune round (teacher generation, annotation and training) cost a further $4.00 (≈ ¥637).
The weights and code are released under Apache-2.0.
tiny-voca-ja is a derivative work of tiny-voca by Muhammed Nazeem, used under Apache-2.0. The architecture, training recipe and code base are his; the Japanese front end, corpus, annotation pipeline and training run are this project's. If you use this model, please credit both.
@software{fujimoto2026tinyvocaja,
author = {Fujimoto, Satoshi},
title = {tiny-voca-ja: a 4.5M-parameter two-voice Japanese text-to-speech model},
year = {2026},
url = {https://huggingface.co/gachimoto/tiny-voca-ja}
}
Original model:
@software{nazeem2026tinyvoca,
author = {Nazeem, Muhammed},
title = {tiny-voca: a 4.5M-parameter two-voice English text-to-speech model},
year = {2026},
url = {https://huggingface.co/n4ze3m/tiny-voca}
}
Author: gachimoto (Satoshi Fujimoto).
The design builds on FastSpeech 2 (explicit duration and prosody prediction), DDSP and the neural homomorphic vocoder (source-filter synthesis with predicted filters), and HiFi-GAN and UnivNet (discriminators). No code or weights from these projects are used. The Japanese alignment uses prj-beatrice/japanese-hubert-base-phoneme-ctc (Apache-2.0) and pyopenjtalk-plus.
onnx/model.onnx 21 MBtiny-voca-ja.pt 18 MBvocoder/gen_final.pt 4 MBvocoder/gen_final.onnx 4 MBn4ze3m/tiny-voca