Configuration Parsing Warning:Config file tokenizer_config.json cannot be fetched (too big)

MoiraTTSAir (female voice)

👩‍💻 Chat with Moira locally on your phone https://play.google.com/store/apps/details?id=com.moiraai.voice

Female-voice Greek fine-tune of NeuTTS-Air for the MoiraAI voice stack: a single-voice speech language model that turns text into NeuCodec audio tokens in Moira's voice. It is fine-tuned to produce this one specific voice — not to clone arbitrary voices.

Built off the open NeuTTS LM + codec design — small on-device backbone, 50 Hz single-codebook neural codec, GGUF-friendly for phones and laptops.

Samples

Reference transcript (samples/greek.txt): "Ο βροντόσαυρος έψαχνε τροφή στην αυλή."

clip audio
Reference voice
Generated example

Model details

  • Voice: female, single voice — Moira. Fine-tuned to speak as Moira; always generate with a reference clip of Moira's voice.
  • Backbone: Qwen2-family causal LM (0.5B scale; NeuTTS-Air builds on Qwen2.5-0.5B) — 24 layers, hidden size 896, 14 attention heads / 2 KV heads, bfloat16.
  • Vocabulary: 217,652 tokens — Qwen2 BPE plus added audio tokens (added_tokens.json).
  • Codec: neuphonic/neucodec (50 Hz, single codebook) — this repo holds the LM half only.
  • Input: phonemised text (handled automatically by the neutts package via espeak-ng).
  • Operating window: 2048 tokens ≈ ~30 s of audio including the voice prompt (config allows up to 32768 positions).
  • Training: fine-tuned for 2 epochs (1,442 steps) — loss curve in trainer_state.json.
  • Sampling defaults: temperature 0.7, top-k 20, top-p 0.8, repetition penalty 1.1 (generation_config.json).
  • Base project: neuphonic/neutts (NeuTTS-Air is Apache 2.0).

Usage

Requires espeak-ng on your system (install guide). This is a public repo — no login needed. Note: the neuphonic/neucodec codec repo is auto-gated, so accept access once while logged in; the backbone itself needs nothing.

pip install neutts soundfile
from neutts import NeuTTS
import soundfile as sf

tts = NeuTTS(
    backbone_repo="moiralabs/MoiraTTSAir",
    backbone_device="cpu",
    codec_repo="neuphonic/neucodec",
    codec_device="cpu",
    language="el",  # required: non-Neuphonic backbones raise without it
)

input_text = "Γεια σου! Χαίρομαι που σε γνωρίζω."
ref_text = open("samples/greek.txt").read().strip()   # transcript of Moira's ref clip
ref_codes = tts.encode_reference("samples/greek.wav") # short clip of Moira's voice

wav = tts.infer(input_text, ref_codes, ref_text)
sf.write("out.wav", wav, 24000)

For on-device/quantised inference, convert the backbone to GGUF and pass the .gguf repo as backbone_repo with llama-cpp-python installed (pip install "neutts[llama]") — same pattern as the upstream neutts-air-q4-gguf / neutts-air-q8-gguf releases.

Other inference paths (see examples/ in neuphonic/neutts):

  • Pre-encoded references: encode_reference() can be run once and the codes saved with torch.save(ref_codes, "ref.pt") to skip re-encoding on every run.
  • ONNX decoder: pass codec_repo="neuphonic/neucodec-onnx-decoder" (needs onnxruntime and a pre-encoded .pt reference) for lower-latency decoding without the codec encoder.
  • Streaming: tts.infer_stream(...) yields audio chunks as they generate, but only works with GGUF backbones (needs pyaudio).

Notes

  • CPU-only machines: plain pip install neutts pulls CUDA torch plus gigabytes of NVIDIA deps. Skip them by installing the CPU builds first (keep their versions matched to each other): pip install torch torchaudio torchvision --index-url https://download.pytorch.org/whl/cpu

  • torchao vs torchtune: current torchao removed NF4Tensor, which torchtune 0.6.1 still imports (No module named 'torchao.dtypes.nf4tensor'). Until upstream agrees, stub it — those names only serve quantization helpers inference never calls:

    # torchao/dtypes/nf4tensor.py
    class NF4Tensor: pass
    def linear_nf4(*a, **k): raise RuntimeError("nf4 shim")
    def to_nf4(*a, **k): raise RuntimeError("nf4 shim")
    def implements(*a, **k): return lambda f: f
    

Files

  • model.safetensors — backbone weights (Git LFS)
  • config.json, generation_config.json — model and sampling config
  • vocab.json, merges.txt, added_tokens.json — tokenizer
  • trainer_state.json, training_args.bin — training provenance
  • chat_template.jinja — Qwen chat template (kept from the base model)
  • samples/greek.wav, samples/greek.txt — Moira reference clip + transcript
  • samples/greek_output.wav — example output in Moira's voice

Limitations

  • Tuned for short Greek utterances.
  • Research-grade voice: expect artifacts on long or out-of-domain text.
  • Audio generated through the NeuTTS pipeline carries a watermark.

Disclaimer

Genos and Moira are fictional AI personas created for the MoiraAI project's own products. Do not use these models to impersonate any real individual or to create misleading recordings, robocalls, scams, or disinformation. You are responsible for complying with all applicable laws on synthetic media and voice rights. We do not condone any misuse of these models.

Acknowledgments

Special thanks to Neuphonic — pioneers of the open text-to-speech space — for releasing NeuTTS and NeuCodec as open source. This model would not exist without their work.

Downloads last month
55
Safetensors
Model size
0.6B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Collection including moiralabs/MoiraTTSAir