Upload folder using huggingface_hub

08329e6 verified 4 days ago

4.22 kB

license: other
license_name: cc-by-nc-4.0-and-apache-2.0
license_link: https://huggingface.co/afrispeech/Akiti-TTS
language:
  - tw
library_name: phoonnx
pipeline_tag: text-to-speech
tags:
  - onnx
  - tts
  - twi
  - akan
  - neutts
  - neucodec
base_model:
  - afrispeech/Akiti-TTS
  - neuphonic/neucodec-onnx-decoder-int8

phoonnx-neutts — Akiti-TTS (Asante Twi), ONNX

ONNX conversion of afrispeech/Akiti-TTS, an Asante Twi text-to-speech model, packaged for phoonnx. This repository holds converted weights only — no new training was done.

Akiti-TTS is a LoRA fine-tune of pnnbao-ump/VieNeu-TTS-0.3B, which comes from Neuphonic's NeuTTS Air family: a Qwen3 causal LM that emits NeuCodec audio tokens.

Files (`akiti-twi-onnx/`)

File	What it is
`neutts_lm.onnx` + `neutts_lm.onnx.data`	Qwen3 backbone, fp32, KV-cached. The graph needs its `.data` sidecar next to it.
`neutts_lm_int8.onnx`	The same graph, dynamically quantized to int8. Smaller, but measurably lower quality — it puts the end-of-speech token in the top 5 at the first step, where fp32 does not.
`neucodec_decoder.onnx`	Copy of neuphonic/neucodec-onnx-decoder-int8 (Apache-2.0), unmodified.
`tokenizer.json`	The checkpoint's own BPE, copied from upstream.
`voices.json`	The nine voice presets, copied from AfriSpeech/akiti-tts (MIT).
`neutts_onnx_meta.json`	Architecture summary written by the export script.

Graph contract

neutts_lm.onnx serves prefill and decode: the same graph with a different past length.

inputs   input_ids       int64 [1, S]         prompt tokens, or 1 token per step
         attention_mask  int64 [1, P + S]     ones over past and current tokens
         position_ids    int64 [1, S]         absolute positions, P .. P+S-1
         past_key_<i>    fp32  [1, 4, P, 64]  i in 0..27
         past_value_<i>  fp32  [1, 4, P, 64]
outputs  logits          fp32  [1, 66938]     last position only
         present_key_<i> / present_value_<i>  fp32 [1, 4, P + S, 64]

neucodec_decoder.onnx takes codes int32 [1, 1, N] and returns audio float32 [1, 1, 480 * (N - 1)] at 24 kHz — 50 codec tokens per second of audio.

Prompt format

The LM is prompted with phonemes, not letters. Text is phonemized by espeak-ng using the lfn (Lingua Franca Nova) voice, which is what the checkpoint was trained with; espeak-ng has no Twi voice, and lfn's five-vowel orthography reads Twi spelling closely.

<|TEXT_PROMPT_START|>{reference phones} {target phones}<|TEXT_PROMPT_END|>
<|SPEECH_GENERATION_START|>{<|speech_c|> for c in reference codes}

Generation continues that speech-token run until <|SPEECH_GENERATION_END|> or EOS.

Fidelity

Exported with scripts/conversion/neutts/export_neutts_onnx.py in phoonnx. Against the torch model on a fixed prompt, max absolute logit difference:

	fp32 ONNX
prefill	4.10e-05
decode (8 steps)	2.96e-05

The fp32 graph's top-5 next tokens on a real prompt are identical to torch's, in the same order and to two decimal places.

Licensing

The pieces carry different licenses, and one of them is inconsistent upstream. Read this before using the model.

afrispeech/Akiti-TTS weights — the Hugging Face model card declares CC BY-NC 4.0 (non-commercial). The GitHub README instead states the weights are Apache-2.0. These two statements disagree. This mirror records both and resolves neither; treat the stricter of the two (non-commercial) as binding until AfriSpeech clarifies. The GitHub repository's code is MIT, which is not in dispute.
neucodec_decoder.onnx — Apache-2.0, from Neuphonic.
voices.json — from the MIT-licensed AfriSpeech/akiti-tts repository.

Attribution: AfriSpeech / Ghana NLP (Akiti-TTS), pnnbao-ump (VieNeu-TTS-0.3B), Neuphonic (NeuTTS Air, NeuCodec).