dida-80b/kokoro-german-hui-multispeaker-base

Kokoro German — HUI Multispeaker Base (Stage 1)

German multispeaker Stage 1 base model built on Kokoro-82M, trained on the HUI Audio Corpus German (CC0, 51 speakers).

This is a base model, not a voice. Use it as the foundation for Stage 2 fine-tuning with your own speaker data. Compared with a single-speaker Stage 1 base, a multispeaker base is less tied to one speaker identity, but it is not speaker-neutral in any strict sense.

Training repo: semidark/kokoro-deutsch · Discussion: Issue #9 · Dataset: dida-80b/hui-german-51speakers

---

License Note

This checkpoint is built on top of Kokoro-82M, which is released under apache-2.0.

The training dataset used here, hui-german-51speakers, is cc0-1.0.

---

Dataset

Corpus Source HUI Audio Corpus German License CC0 Raw speakers 122 After filtering 51 (24m / 27f) Raw speaker duration strongly imbalanced Total effective ~51h

Train / Val Split Train samples 20,495 Val samples 418 Clip duration 1–20s Sample rate 24 kHz Format WAV (mono)

Quality Filter Check Threshold Min RMS −42 dB Max clipping 0.1% Max silence 50% Min per speaker 5 min Manual review ✓ all 51

Duration cap: 60 min/speaker prevents dominant speakers from biasing the model (Bernd_Ungerer alone contributed 81h raw).
Weighted sampling: Small speakers are duplicated so all 51 appear equally often in batches — speaker embeddings receive identical gradient update rates throughout training.

---

Training

Setup Base model kokoro_base.pth GPU NVIDIA A40 (48 GB) Batch size 8 Epochs 10 Workers 7

Avg. Mel Loss per Epoch Epoch Mel Loss (avg) Δ 1 0.5826 — 2 0.4131 −29% 3 0.3743 −9% 4 0.3577 −4% 5 0.3483 −3% 6 0.3414 −2% 7 0.3364 −1% 8 0.3321 −1% 9 0.3288 −1% 10 0.3264 −44% total

Final Losses (Ep 10) Loss Value Mel Loss 0.333 Disc Loss 3.983 Mono Loss 0.030 S2S Loss 0.523 SLM Loss 1.422 Mono Loss staying low throughout training indicates stable attention alignment — the most common Stage 1 failure mode did not occur.

---

Files

File	Description
first_stage.pth	Final Stage 1 checkpoint (Epoch 10) — use this for Stage 2 fine-tuning

---

Audio Samples — Epoch 10

7 phonetically diverse test sentences synthesized from the final checkpoint.
Note: this is a multispeaker base checkpoint, not a single-speaker voice model.

#	Text	Audio
1	Schön, dass du da bist. Die Bücher liegen auf dem großen Tisch.
2	Ich mache mich auf den Weg nach Aachen, um auch nachts wach zu sein.
3	Er aß die Maße in der Straße, aber das Maß war voll.
4	Zwei weiße Zwerge zwängen sich zwischen zwei Zweige.
5	Ein Pfau pflegt seine Federn an der Pfütze.
6	Warum hast du das getan? Das ist ja unglaublich!
7	Das kostet genau einhundertdreiundzwanzig Millionen Euro.