magpie_tts_multilingual_357m-he

<img src="assets/banner.jpg" alt="MagpieTTS Hebrew" width="100%">

Hebrew fine-tune of nvidia/magpie_tts_multilingual_357m. Input is IPA, not Hebrew letters — there is no G2P in the model, you pass phonemes.

Training code: github.com/maxmelichov/magpie_tts_multilingual_357m-he

Model: notmax123/magpie_tts_multilingual_357m-he

Results

IPA WER on 1,525 held-out utterances of Phonikud/ILSpeech, scored by notmax123/whisper-he-ipa.

<img src="ipa_wer_chart.png" alt="IPA WER for notmax123/magpie_tts_multilingual_357m-he" width="100%">

Limitations

No voice cloning — voice selection instead. The released base checkpoint has the context_encoder weights removed entirely (0 tensors in the file), replaced by 5 baked speaker embeddings — NVIDIA dropped zero-shot cloning for security reasons. context_audio and context_text are both ignored; the reference clip has no effect on the output voice. We extended the baked table with 15 trained Hebrew voices, selected by --speaker-index (see Usage). Self-similarity per voice is 0.74-0.80 (real-speaker scale ~0.80), cross-voice 0.11-0.27 (different speakers ~0.00) — each index is a stable, distinct voice. The original 5 released voices (indices 0-4) are restored bit-exact after training and are unaffected.

Loanword phonemes are missing. RenikudPlus emits w and the geresh set ג׳ ז׳ צ׳ -> dʒ ʒ tʃ, and of those only tʃ works — the tokenizer is character-level, so tʃ is just t+ʃ, both trained. w, ʒ and dʒ come out wrong (wiskˈi -> vˈiskij, hadʒˈip -> hadˈip). scripts/extend_vocab.py adds w/ʒ to a trained checkpoint and seeds them from the base model's Spanish vocabulary, but seeding alone does not transfer pronunciation — unlike the other 26 symbols, these have no gradient updates behind them.

Training cannot fix this from the current corpus: it contains 7 w and 1 ʒ in 1.3M utterances, and re-phonemizing the Hebrew source with RenikudPlus yields ~0 ʒ/dʒ and ~262 w across 788k rows — the text is formal Hebrew with no geresh marks at all. It needs targeted loanword recordings, not more of the same data.

Training data is synthetic (TTS output verified by ASR, not studio recordings), which caps naturalness near the teacher systems'.

Other languages are intact. English WER measured identical to the base model (4.2%, same transcripts) — training was masked to the 33 new Hebrew embedding rows only.

Usage

Inference

The model takes IPA phonemes, not Hebrew letters — there is no G2P inside it. Inference lives in the training-code repo.

git clone https://github.com/maxmelichov/magpie_tts_multilingual_357m-he
cd magpie_tts_multilingual_357m-he
bash scripts/setup.sh   # venv + NeMo (needed by scripts/infer_hebrew.py)
venv/bin/hf download notmax123/magpie_tts_multilingual_357m-he --local-dir checkpoints/hebrew
# magpie_hebrew.ckpt, hparams.yaml, tokenizer assets under assets/

A local release/ dir (model.ckpt + hparams.yaml, gitignored) is the same portable bundle — point --checkpoint / --hparams at it instead of the Hub files.

Phonemize with RenikudPlus (tools/renikud/). Strip nikud first — it predicts the diacritics itself and mangles pre-pointed input.

venv/bin/hf download notmax123/RenikudPlus model_int8.onnx --local-dir tools/renikud
venv/bin/pip install onnxruntime
venv/bin/python -c "
import sys; sys.path.insert(0, 'tools/renikud')
from renikud_onnx import G2P
print(G2P('tools/renikud/model_int8.onnx').phonemize('שלום, מה שלומך?'))
"
# ʃalˈom, mˈa ʃlomχˈa?

Then synthesize:

venv/bin/python scripts/infer_hebrew.py \
  --checkpoint checkpoints/hebrew/magpie_hebrew.ckpt \
  --hparams checkpoints/hebrew/hparams.yaml \
  --speaker-index 13 \
  --text "ʃalˈom, mˈa ʃlomχˈa?" --context-audio <any 10s+ wav> --out-dir outputs/

--speaker-index is the only voice control (see Limitations); --context-audio is required by the plumbing but does not change the voice (the context_encoder was removed). The clip must be >= 10s. --gpu selects the CUDA device (script default is 1). Indices 0-4 are the released NVIDIA voices (Aria, Jason, John, Leo, Sofia), unaffected by this fine-tune. 5-19 are the Hebrew voices added here:

Supported IPA symbols (27):

a b d e f h i j k l m n o p r s t u v z ɡ ʁ ʃ ʔ χ ˈ   plus  , . ? !

Training

bash scripts/setup.sh                          # venv + NeMo + base checkpoint
venv/bin/python scripts/build_manifests.py     # manifests from the Hebrew CSVs
bash scripts/train_hebrew.sh +num_new_speakers=15   # train (both GPUs); 0 for single-voice instead

venv/bin/python scripts/merge_lora.py \
  --ckpt experiments/Magpie-TTS/checkpoints/<best>.ckpt --out merged.ckpt

venv/bin/python scripts/infer_hebrew.py \
  --checkpoint merged.ckpt --hparams experiments/Magpie-TTS/version_0/hparams.yaml \
  --speaker-index 13 \
  --text "ʃalˈom, mˈa ʃlomχˈa?" --context-audio <any 10s+ wav> --out-dir outputs/

Evaluation

venv/bin/python scripts/build_ilspeech_eval.py --split all --out-name eval_full.json
venv/bin/python scripts/infer_hebrew.py --manifest data/ilspeech/eval/eval_full.json ...
venv/bin/python scripts/score_ilspeech.py --pred-dir outputs/<run>/audio/repeat_0

scripts/extend_vocab.py adds IPA symbols to a trained checkpoint without retraining, and scripts/rephonemize_renikud.py re-phonemizes the corpus with RenikudPlus (strip nikud first — it predicts the diacritics itself and mangles pre-pointed input).

scripts/diagnose_speaker_conditioning.py checks whether the reference clip changes the output at all, printing same-speaker and different-speaker anchors so the similarity number has a scale.

How it works

LoRA (r=16, α=32) on attention projections — 78 adapters, 4.7M of 375M params trainable, merged into the base weights for release. 80k steps on 2× RTX 5090.

The baked speaker table (see Limitations) is extended from 5 to 20 rows and its new rows made trainable; a gradient mask keeps the original 5 frozen during training, and merge_lora.py restores them bit-exact from the base checkpoint regardless. Set +num_new_speakers=0 to skip this and train a single-voice model instead — measurably better WER, no voice selection (see Results).

Hebrew is a 16th tokenizer appended to the checkpoint's 15, so pretrained token IDs keep their offsets. Four things that were not obvious and cost real debugging time:

• BOS/EOS move when you append tokens. The table is [vocab | new | BOS | EOS], so copying the checkpoint as one contiguous prefix drops the pretrained BOS/EOS onto the first two new tokens and leaves the real ones random. Copy vocabulary and special rows separately.
• 26 of 27 Hebrew IPA symbols already exist in the base model's Spanish/Portuguese/Hindi phoneme vocabularies. Seed the new rows from those instead of from noise; only ʔ is genuinely new.
• The shared embedding table spans all 16 languages. A gradient mask keeps training to the new rows, otherwise Hebrew drags the other languages with it.
• The released checkpoint uses frame stacking (factor 2), so the stock config will not load into it — rebuild the model config from the checkpoint's own model_config.yaml — and context clips must be >= 10 s.

Data: ~285k utterances / 292 hours across 14 Hebrew speakers, filtered to wer_score == 0.0 (exact ASR agreement). That strict filter beat <= 0.2 by 0.36% WER (95% CI [-0.65, -0.06]) — real, but small, and it costs 63% of the data.

License

Derivative of nvidia/magpie_tts_multilingual_357m under the NVIDIA Open Model License. ILSpeech is non-commercial and was used for evaluation only — none of it is in the weights.