You can deploy this model and start earning money today!

ThaoTran7/
vocence-voxcpm2

Name: vocence-voxcpm2
Availability: InStock

Views

No views yet

Vocence VoxCPM2 miner (reference HF repo layout)

A production-ready Vocence PromptTTS miner built on OpenBMB VoxCPM2 (model card).

Backbone: VoxCPM2 (2B params, MiniCPM-4 + AudioVAE V2)
License: Apache-2.0 — safe for commercial subnet mining
Output: mono float32 PCM at 48 kHz
VRAM: ~8 GB live (fits on a 16 GB RTX 4000 with headroom for KV cache)
Mode used: Voice Design — the model accepts a parenthesised natural-language description of the speaker, no reference audio required. Vocence's instruction string maps almost 1:1 to this mode.

Why VoxCPM2 over Parler-TTS / Fish S2-Pro for Vocence

Axis	Parler v1	Fish S2-Pro	VoxCPM2
License	Apache-2.0	Research-only (NOT commercial)	Apache-2.0
Description-conditioned synthesis	Yes (English-centric)	No (voice-cloning)	Yes, native Voice Design
Sample rate	44.1 kHz	44.1 kHz	48 kHz (built-in super-res)
Training data	~50 k hrs	~1 M hrs	2 M+ hrs
Languages	English-heavy	Multilingual	30 languages natively
InstructTTSEval	Strong	N/A	SOTA per model card
RTF on RTX 4000 16 GB	~1.0–1.5	Tight	~0.5–0.6
Built-in LoRA fine-tune	Community	Community	Official scripts

For the Vocence task — text + instruction → audio scored on script (0.30) + naturalness (0.15) + 7 trait fields (0.55) — VoxCPM2 wins on every axis that moves the needle.

Files

File	Purpose
`miner.py`	Engine. `class Miner` with `__init__(path_hf_repo)`, `warmup()`, `generate_wav(instruction, text)`. Translates Vocence's pipe-delimited instruction into a VoxCPM2 Voice Design prompt and synthesizes 48 kHz audio.
`chute_config.yml`	Image (CUDA torch + voxcpm + pre-baked VoxCPM2 weights), NodeSelector (16 GB GPU), Chute settings.
`vocence_config.yaml`	Runtime knobs: `cfg_value`, `inference_timesteps`, duration cap, request limits.

Instruction → Voice Design mapping

Vocence sends:

text

gender: female | pitch: high | speed: normal | age_group: adult |
emotion: happy | tone: warm | accent: us

miner.py expands this into a single VoxCPM2 description and prepends it to the text in parentheses (the model's documented Voice Design syntax):

text

(A middle-aged adult female speaker with a bright, high-pitched, warm voice,
 sounding happy and cheerful, speaking at a natural, moderate pace, with a
 clear American English accent, clean studio recording, no background noise,
 no reverb)Hello, world.

Two reasons this beats a literal (gender: female | pitch: high | ...)Hello:

VoxCPM2 was pre-trained on descriptive English speaker captions, not key-value strings, so descriptive phrasing lands inside its training distribution.
The "clean studio recording, no background noise" cues are a small but reliable lift on the GPT-4o pairwise naturalness call (15% of the score).

The mapping lives in pure dictionaries (_GENDER, _PITCH, _SPEED, _AGE, _EMOTION, _TONE, _ACCENT) at the top of miner.py — change the wording there, re-run the local harness, ship.

Local iteration loop

bash

# from the vocence repo root
uv run python scripts/local_miner_eval.py \
    --repo miner_sample/example_repo \
    --out  /tmp/vocence_local_eval \
    --num  16

# Optional: enable GPT-4o-audio scoring (matches docs/scoring.md exactly)
OPENAI_API_KEY=sk-... uv run python scripts/local_miner_eval.py \
    --repo miner_sample/example_repo --num 32 --judge

Outputs WAV files plus a report.json with per-case timings, RTF, and (with --judge) Vocence-shaped element scores and a win-rate aggregate.

Deploy

Important: the built-in vocence miner push CLI command currently generates a non-canonical chute (/synthesize endpoint, TTSArgs schema, hard-coded AutoModelForTextToSpeech) and would fail the owner's wrapper-integrity hash check. Use the renderer below instead — it produces a chute that is byte-identical to the canonical template except for the four approved variables.

Push this directory to your own Hugging Face repo. The repo must contain miner.py, chute_config.yml, and (optionally) vocence_config.yaml. Note the commit SHA you want to deploy.
Render the canonical wrapper with your four approved variables:

bash
1uv run python scripts/render_chute.py \ 2 --repo you/your-vocence-miner \ 3 --revision <40-char-commit-sha> \ 4 --chute-user your-chutes-user \ 5 --chute-name vocence-voxcpm2-001 \ 6 --out chute_deploy.py

The renderer validates that:
- --repo matches user/name
- --revision is a hex SHA (no branch names — branches drift, hashes don't)
- --chute-name contains vocence (required by owner validation)
Build and deploy on Chutes:

bash
1chutes build chute_deploy:chute --wait 2chutes deploy chute_deploy:chute --accept-fee

The first build will take a while because chute_config.yml pre-bakes the VoxCPM2 weights into the image.
Commit on chain:

bash
1uv run vocence miner commit \ 2 --model-name you/your-vocence-miner \ 3 --model-revision <same-sha-as-step-2> \ 4 --chute-id <UUID-from-chutes-deploy>

vocence miner commit is the only half of the miner CLI that is correct today and matches what validators expect.

Tuning playbook

Knobs in vocence_config.yaml, in priority order of impact:

Prompt phrasing in the _* maps in miner.py — biggest lever; iterate with the local harness.
cfg_value (default 2.0) — raise to 2.5–3.0 for stronger trait adherence at a small naturalness cost.
inference_timesteps (default 10) — 16–24 marginally cleaner but ~2× slower; only worth it if you have RTF headroom.
LoRA fine-tune (Phase 3) — VoxCPM2 ships official LoRA scripts that work on a single 16 GB GPU. Fine-tune on (rich_description, audio) pairs generated by your own mapper to lock in trait fidelity.

License

This example repo: MIT (matches the Vocence subnet repo). The VoxCPM2 weights themselves are Apache-2.0 from OpenBMB.

ThaoTran7/vocence-voxcpm2