Downloads · 30 days
0
cosyvoice3-rk3576-docker-tts
keelzhang/cosyvoice3-rk3576-docker-tts
cosyvoice3-rk3576-docker-tts is a text-to-speech model from keelzhang. Use it when you need text read aloud. It is set up for cosyvoice3. The card lists the license as apache-2.0.
Board-side RKNN / RKLLM pack for the CosyVoice3 Docker TTS service in keelzhang/cosyvoice3-rk3576 (docker/tts).
Access
Public
Updated Aug 26, 2026
Repo size
1.9 GB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.rknn780 MB · 42%
At a glance
- Task
- Text-to-Speech
- Library
- cosyvoice3
- License
- apache-2.0
- Access
- Public
- Created
- Aug 26, 2026
- Updated
- Aug 26, 2026
- SHA
- 7abfd97e
Model card
From the Hugging Face model README
CosyVoice3 RK3576 — Docker TTS weights
Board-side RKNN / RKLLM pack for the CosyVoice3 Docker TTS service in
keelzhang/cosyvoice3-rk3576
(docker/tts).
Code and docs stay on GitHub; this repo is weights only.
What this is for
Mount (or download into) /opt/cosyvoice3 on an RK3576 board, then run the
thin FastAPI container from the GitHub repo. The service:
- loads RKLLM + RKNN once and keeps handles resident
- runs a discarded warmup utterance before
warmed=true - exposes
GET /health,GET /ready,POST /v1/tts→ WAV - caps utterances at 20 seconds
Ops guide: docs/BOARD_TTS.md
Download
# huggingface_hub / huggingface-cli
huggingface-cli download keelzhang/cosyvoice3-rk3576-docker-tts \
--local-dir /opt/cosyvoice3
Or with Python:
from huggingface_hub import snapshot_download
snapshot_download(
"keelzhang/cosyvoice3-rk3576-docker-tts",
local_dir="/opt/cosyvoice3",
)
Directory layout is flat (same as the board mount expected by docker/tts).
File map
| Path | Role |
|---|---|
qwen2_body_w8a8_embedcalib_c3_ctx2048_rk3576.rkllm | RKLLM speech-token body |
embed_tokens.npy | Text / LLM embedding table |
speech_embedding.npy | Speech token embedding table |
speech_head_fp16.rknn | Speech head (logits) |
flow_encoder_mu_fp16_tok384_rk3576.rknn | Flow encoder (tok pad 384) |
flow_estimator_fp16_seq100_rk3576.rknn | Flow estimator (seq 100) |
hift_f0_fp16_mel100_rk3576.rknn | HiFT F0 (mel 100) |
hift_decoder_fp16_mel100_rk3576.rknn | HiFT decoder (mel 100) |
hift_m_source_linear.npz | HiFT source linear weights |
tokenizer/ | Text tokenizer (BlankEN-compatible tree) |
voices/xilian/prompt_speech_tokens.npy | Offline prompt speech tokens |
voices/xilian/spks.npy | Speaker embedding |
README.md | This file |
Not included (optional / other workflows):
- PC
flow.pt/ full Fun-CosyVoice3 pretrained tree - Host product student ckpt
flow_student_stream_v2.pt(acoustic RTF benches) - Extra HiFT mel buckets / H5-L incremental stack beyond mel100 oneshot Docker path
- Runtime
.py(ship with the GitHub repo; keep them on the board next to weights)
Board runtime extras
Besides this pack you still need on the device:
- GitHub
docker/ttscompose + image - Host Python site-packages mount (see compose
PYTHONPATH) librknnrt.so/librkllmrt.so- Flat board helper modules under
/opt/cosyvoice3(fromruntime/in the git repo), e.g.board_infer_demo.py,rkllm_ctypes.py, …
License & attribution
- Packaging / docs for this weight layout: Apache-2.0 (same as the GitHub code repo).
- Underlying CosyVoice / Fun-CosyVoice3 / Qwen-derived artifacts remain under upstream licenses — redistribute and use accordingly.
- Reference voice: only token / speaker embeddings are shipped here (no prompt WAV).
- Task
- Text-to-Speech
- Library
- cosyvoice3
- License
- apache-2.0
- Created
- Aug 26, 2026
- Updated
- Aug 26, 2026