Pocket TTS — Greek (6 layers)
Greek and Cypriot Greek text-to-speech for Pocket TTS, running faster than real time on a CPU.
- Trained on ~2,400 hours of Greek speech from 3,900+ speakers, balanced between female and male voices.
- As intelligible as real speech: Whisper transcribes it as accurately as real recordings of the same sentences.
- Voice cloning from a 10–15 s sample, plus eight curated voices, two of them Cypriot.
- ~110 M parameters, 3× faster than real time on a laptop CPU, with streaming output.
- Own Greek tokenizer, trained on correctly punctuated Greek text.
Voices
Every clip below was generated by this model, not recorded.
| voice | description | sample |
|---|---|---|
| Eleni | female, young adult, calm and warm | |
| Athina | female, young adult, clear and professional | |
| Katerina | female, adult, upbeat and expressive | |
| Sophia | female, young adult, Cypriot, bright and lively | |
| Spyros | male, adult, deep, calm and warm | |
| Kostas | male, young adult, clear and professional | |
| Nikos | male, adult, deep, upbeat and friendly | |
| Giorgos | male, mature, Cypriot, expressive |
Use a voice by passing its file, e.g. --voice hf://myned-ai/pocket-tts-greek/voices/eleni.wav.
Descriptions are approximate. voices/voices.md lists the source recording and the cleaning
applied to each voice. The named English voices (alba, …) do not work with this model; any other
clean recording of a speaker does.
Usage
uvx pocket-tts generate \
--config hf://myned-ai/pocket-tts-greek/greek.yaml \
--voice hf://myned-ai/pocket-tts-greek/voices/kostas.wav \
--text "Καλημέρα σας! Θέλετε να κλείσουμε ένα ραντεβού για αύριο;"
from pocket_tts import TTSModel
model = TTSModel.load_model(config="hf://myned-ai/pocket-tts-greek/greek.yaml")
voice = model.get_state_for_audio_prompt("hf://myned-ai/pocket-tts-greek/voices/eleni.wav")
audio = model.generate_audio(voice, "Γεια σας, πώς μπορώ να σας βοηθήσω;")
- Questions end with
;, the Greek question mark. The config keeps it; do not convert it. - Write numbers as words for best results ("στις δέκα και μισή", not "στις 10:30").
- Default temperature is 0.3; 0.5–0.7 gives livelier but less even delivery.
- For your own voice: 10–15 s of one continuous, clean take, ideally with the speaking style you want. The model copies the recording's quality, room and pace.
Training
| data | 2,395 h of Greek speech, 766k utterances, 3,900+ speakers, 50 % female; 115 h Cypriot. Mostly Greek podcasts (automatically transcribed and aligned) plus Mozilla Common Voice (el, el-CY). Sources and timestamps: myned-ai/pocket-tts-greek-data |
| tokenizer | SentencePiece, 4,000 tokens, trained on the transcripts |
| teacher | 24 layers, fine-tuned from english_2026-04_24l with a fresh text embedding, 300k steps |
| student | 6 layers, depth-distilled from the teacher (guidance baked in), 250k steps |
| punctuation | The ASR marked questions with the Latin ? instead of the Greek question mark ;. We converted them to the correct Greek form before training. |
| hardware | one GH200, ~23 h in total |
Trained with the Pocket TTS training code.
Evaluation
80 Greek sentences from 10 Common Voice speakers never seen in training. The voice prompt is a different recording of the same speaker. Whisper large-v3 is the judge, at temperature 0.3.
| this model | real recordings | |
|---|---|---|
| word error rate | 11.2 % (one Whisper hallucination excluded) | 12.5 % |
| speaker similarity (WavLM) | 0.912 | 0.937 (two real clips) |
| UTMOS | 2.98 | 2.46 |
Whisper misreads 12.5 % of words even in real Greek speech, and the test sentences are largely literary, so absolute error rates here are not comparable with English models.
Across the 8 voices × 59 agent-style sentences, 87 % of clips were read cleanly.
Limitations
- Very short replies ("Εντάξει.", "Χαίρετε.") are the weakest case: about 1 in 5 comes out unclear. Prefer "Εντάξει, το κανονίζω." to a bare "Εντάξει.".
- Questions rise at the end more weakly than in natural speech.
- Rare words outside everyday conversation ("πλησιέστερο") are sometimes misread.
- Cypriot could not be scored automatically, because Whisper misreads real Cypriot speech about 46 % of the time. It has been judged by ear only.
- Voice cloning: only clone voices you have permission to use.
License
Weights: CC BY 4.0, the same as the base model kyutai/pocket-tts.
Commercial use is allowed with attribution. The bundled voices are cut from Mozilla Common Voice
(CC0) recordings; please do not attempt to identify the contributors. As upstream asks: do not use
this model to impersonate people or to present generated audio as genuine recordings.