ShinyUser/vocence-miner04

vocence_miner_v3

A reliability-and-naturalness pass over the prompt-driven Qwen3-TTS-12Hz-1.7B-VoiceDesign backbone. v3 ships two changes that matter at inference time:

1. Full-sentence generation. Earlier checkpoints would sometimes render only the first clause of a longer input — the rest of the sentence would be cut off, dropped, or replaced with silence. v3 generates the entire input from start to end, including longer sentences with intermediate clauses, em-dashes, and parenthetical asides.

2. More natural delivery. Across the same prompt set, v3 produces audibly smoother prosody — fewer flat reads on neutral prompts, less "narrated" surface on short utterances, and more believable breath placement on persona reads.

Everything else stays the same: free-form English instruct, 24 kHz mono output, single-call inference, no reference audio.

---

Use it

pip install qwen-tts transformers torch soundfile

from qwen_tts import Qwen3TTSModel
import soundfile as sf

m = Qwen3TTSModel.from_pretrained("magma90909/vocence_miner_v3")

wavs, sr = m.generate_voice_design(
    text="When I got home, the lights were on, the back door was wide open, and somebody had left tea brewing on the kitchen counter.",
    instruct="A nervous middle-aged man recounting the moment, slightly hushed, slightly fast.",
    language="english",
)
sf.write("out.wav", wavs[0], sr)

The example deliberately uses a long, multi-clause sentence — the kind that earlier checkpoints would clip mid-read.

---

What instruct understands

Axis	Working values
Gender	male, female
Pitch	deep, low, medium, high, thin
Pace	slow, halting, moderate, brisk, fast
Affect	neutral, happy, sad, angry, fearful, urgent, calm, projected, whispered, sarcastic
Persona	bedtime storyteller, news anchor, sports announcer, stern parent, weary narrator

Lead with gender on emotion-heavy prompts to avoid timbre drift.

---

Caveats

• English only — other languages were not part of this checkpoint's adaptation set.
• Strongly expressive reads (drawn-out sad reads, projected announcer reads) may run slightly less precise on automatic transcription than the base. The trade-off was made deliberately for delivery character.
• CC BY-NC-SA 4.0 — research and non-commercial use only.

---

What's in the repo

• model.safetensors — merged Talker weights
• speech_tokenizer/ — Qwen3 12 Hz audio codec
• tokenizer.json, vocab.json, merges.txt, configs — text-side assets
• miner.py, chute_config.yml, vocence_config.yaml — Vocence engine glue (TEE / pro_6000)
• demo.py — quick smoke test

The Vocence files make this repo deployable on Bittensor SN78 (Vocence) via the canonical Vocence/Chutes wrapper without modification.

---

Local A/B evaluator (subnet-like)

Use eval_ab.py to compare two checkpoints with Vocence-like weighting and pass threshold:

pip install qwen-tts transformers torch soundfile audiojudge
python eval_ab.py \
  --a-source . \
  --b-source <other_model_or_path> \
  --dataset ./eval_prompts.jsonl \
  --openai-key $OPENAI_API_KEY \
  --device cuda:0

Dataset format (.jsonl, one object per line):

{"text":"hello world","instruction":"gender: male | pitch: mid | speed: normal | age_group: adult | emotion: neutral | tone: casual | accent: us","source_audio":"/abs/path/source.wav"}

source_audio is optional, but recommended to include naturalness scoring like subnet validation.