You can deploy this model and start earning money today!

dailouis58/
vocence_miner

Name: vocence_miner
Availability: InStock

safetensors

qwen3_tts

Views

No views yet

Vocence miner — Qwen3-TTS + narrator-prompt + broadcast-DSP

Deployable HF-repo layout and miner.py for a Vocence PromptTTS miner. Strategy:

Weights: a Qwen3-TTS-12Hz-1.7B-VoiceDesign fine-tune (snapshotted into the HF repo itself so the chute does not re-download anything at startup).
Instruction rewriting: the 8-field validator enum string is expanded into a rich audiobook-narrator description; Qwen responds much better to prose.
Text normalization: abbreviations (Mr., Dr., etc.), smart quotes and dashes are normalized before synthesis — directly lowers WER.
Sampling: temperature=0.75, subtalker_temperature=0.70, top_p=0.90, repetition_penalty=1.05. Conservative vs. Qwen defaults, keeps the output stable run-to-run.
Post-DSP: 80 Hz HPF → 14 kHz LPF → sub-2 kHz room-tone noise (-60 dB) → LUFS normalize to −20 LUFS → soft tanh saturation. This is what makes the clip read as real studio audio to GPT-4o's pairwise naturalness judge.
Hard node: pro_6000 only, so the output is bit-stable across validator calls — same trick both top miners use.

Files

File	Purpose
`miner.py`	TTS engine (class `Miner` + Vocence contract)
`chute_config.yml`	Chutes Image/NodeSelector/Chute (pinned deps, pro_6000)
`vocence_config.yaml`	Runtime config (dtype, language, limits) read by `miner.py`
`prepare_repo.py`	Helper: pulls source checkpoint, strips leak files, overlays our miner files, perturbs one tensor (duplicate-hash dodge)
`README.md`	This file

Deploy flow

0. Prerequisites

Chutes developer account + API key (CHUTES_API_KEY).
HuggingFace account + write access to a private or public repo.
HuggingFace CLI + git-lfs installed.
Your Bittensor wallet (coldkey + hotkey) already registered on subnet 78.

1. Stage the HF repo locally

bash

cd /root/workspace/vocence
python miner/prepare_repo.py \
    --source macminix/qwen3_voice_design_t1 \
    --revision 2530e9dc \
    --out /tmp/my-vocence-repo

This pulls the full ~4.5 GB of weights, strips eval_results.json, example_inference.py, any checkpoint/training leak files, overlays our miner.py / chute_config.yml / vocence_config.yaml / README.md, and adds a single-scalar perturbation of magnitude 1e-8 to one tensor of the largest .safetensors. That perturbation changes the file's sha256 — which is required for Vocence owner validation to avoid detect_duplicates → duplicate_model. Numerically the perturbation is far below bfloat16 rounding noise so model behaviour is unchanged.

Do not skip the perturbation. detect_duplicates keys off the sha256 of your weight files. If that matches an earlier miner's hash, you are marked invalid and earn zero — regardless of winrate.

2. Push to your HF repo

Replace <your-user>/<your-repo> everywhere below.

bash

huggingface-cli login        # first time only
huggingface-cli upload <your-user>/<your-repo> /tmp/my-vocence-repo . --revision main

Record the commit sha HF prints (you need it in step 4). You can also grab it with:

huggingface-cli api repo_info --repo-id <your-user>/<your-repo> | jq -r .sha

3. Render + build + deploy the chute

From the Vocence monorepo, render the canonical chute template with your four variables:

bash

cd /root/workspace/vocence
python - <<'PY'
from pathlib import Path
from jinja2 import Template

tpl = Path("miner_sample/chute_template/vocence_chute.py.jinja2").read_text()
rendered = Template(tpl).render(
    huggingface_repository_name="<your-user>/<your-repo>",
    huggingface_repository_revision="<hf-commit-sha-from-step-2>",
    chute_username="<your-chutes-username>",
    chute_name="my-vocence-tts-01",   # MUST contain "vocence" (case-insensitive)
)
Path("tmp_chute.py").write_text(rendered)
print("rendered -> tmp_chute.py")
PY

chutes build tmp_chute:chute --wait
chutes deploy tmp_chute:chute --accept-fee

The deploy command prints a chute UUID — that is the chute_id you commit on chain. Remember: the chute name (my-vocence-tts-01 above) is the string the owner checks for the magic word "vocence"; the chute_id is the UUID assigned by Chutes and does not need to contain "vocence".

4. Commit on chain

bash

cd /root/workspace/vocence
uv run vocence miner commit \
    --model-name "<your-user>/<your-repo>" \
    --model-revision "<hf-commit-sha-from-step-2>" \
    --chute-id "<chute-uuid-from-step-3>" \
    --coldkey "<your-coldkey-name>" \
    --hotkey "<your-hotkey-name>"

Wait a block or two and confirm:

uv run vocence get-miners | grep <your-hotkey-prefix>

5. Wait for eligibility

The validator only counts you toward the leaderboard once you have

more than 40 evaluations in each of at least 3 active validators (MIN_EVALS_TO_COMPETE = 40, MIN_VALIDATOR_APPEARANCES_FOR_ELIGIBILITY = 3).

At ~1 eval per validator per 30 min, that is on the order of ~20 hours of uptime before you can appear as a winner. Keep the chute hot the whole time — any failed /speak call means the validator skips adding an eval row for you that round.

Health / debug

Once deployed you can poke the chute directly:

bash

curl -s "https://<chute-slug>.chutes.ai/health" | jq .
curl -s -X POST "https://<chute-slug>.chutes.ai/speak" \
    -H "Authorization: Bearer $CHUTES_API_KEY" \
    -H "Content-Type: application/json" \
    -d '{"instruction":"gender: male | age_group: adult | pitch: mid | speed: normal | emotion: calm | tone: warm | accent: us","text":"The quick brown fox jumps over the lazy dog."}' \
    --output /tmp/out.wav

dailouis58/vocence_miner

Vocence miner — Qwen3-TTS + narrator-prompt + broadcast-DSP

Files

Deploy flow

0. Prerequisites

1. Stage the HF repo locally

2. Push to your HF repo

3. Render + build + deploy the chute

4. Commit on chain

5. Wait for eligibility

Health / debug

dailouis58/
vocence_miner