mlx-community/Irodori-TTS-v4.1-Small-8bit

Use with mlx-audio

Command line

Python

Notes

• Unified v4.1 model: voice cloning, VoiceDesign (caption) and automatic duration prediction in one checkpoint.
• v4.1 differs from v4 only in the duration predictor, which upstream retrained separately with every other parameter frozen. The remaining 683 of 714 tensors are bit-identical to v4.
• The ModernBERT-ja-310m text encoder weights and its tokenizer are bundled, and the Semantic-DACVAE-Japanese-32dim codec ships in dacvae/, so inference needs no additional downloads.
• Reference audio: up to 120s. Passing a list of clips encodes each separately and concatenates them, which matches training better than one long recording.
• Short caption-only prompts are shorter than in v4, but still over-predicted. Measured with the same reference clip and caption:

  Text	Tokens	v4 caption only	v4.1 caption only
  こんにちは。	3	3.64s	2.88s
  今日はいい天気ですね。	5	5.60s	4.72s
  MLXへの移植が完了しました。	7	4.08s	3.84s

  Under about seven tokens the model can still fill the surplus by reading the sentence a second time. This matches the reference PyTorch implementation, which predicts 117.41 frames against MLX's 117.56 for the second row, and produces the same repeat. Pass reference audio, or set duration_scale / seconds.
• Quantized to 8 bits with mlx.nn.quantize (group size 64). Note that this covers more layers than the upstream torchao checkpoints, which keep AdaLN, the projectors and the duration predictor in BF16. Duration predictions stay within one frame (40 ms) of the fp16 model.