Faster Qwen3-TTS Demo

This Space hosts the demo UI for faster-qwen3-tts with streaming audio, TTFA/RTF metrics, voice clone, custom voices, and voice design. It defaults to the GGML/qwentts.cpp backend and includes a Settings toggle to switch between GGML and the Torch CUDA-graph backend for speed comparisons.

Run locally (no Docker)

pip install --index-url https://download.pytorch.org/whl/cu128 torch torchaudio
pip install "faster-qwen3-tts[demo,ggml]==0.3.0" nano-parakeet
python server.py --backend ggml --model Qwen/Qwen3-TTS-12Hz-0.6B-Base
# open http://localhost:7860

Run with Docker

docker build -t faster-qwen3-tts-demo .
docker run --gpus all -p 7860:7860 faster-qwen3-tts-demo

Auth and quotas

Local launches default to an open demo with no login requirement. The Space Docker image sets DEMO_WEB_ONLY=1 and DEMO_REQUIRE_LOGIN=1, and the Space enables Hugging Face OAuth via hf_oauth: true. In that mode, users must sign in before the app UI or API routes are usable. Generation requests are tracked by an HMAC of the user's stable Hugging Face OAuth ID: non-PRO users get DEMO_DAILY_FREE_REQUESTS generations per UTC day, while PRO users are not limited. The raw OAuth ID is not stored. For admin visibility, the database also keeps a usage_users table mapping each HMAC user key to the latest observed Hugging Face username and PRO status.

Usage is stored in a SQLite database inside an attached Hugging Face Storage Bucket. Create a bucket and attach it to the Space as a read-write volume at /data; the app will automatically store usage at /data/faster-qwen3-tts-usage.sqlite3. If the bucket is not mounted, usage falls back to temporary storage and resets when the Space restarts.

When DEMO_WEB_ONLY=1, expensive POST routes require a short-lived token that is injected into the served web UI and sent back by the browser. This blocks casual direct API calls, but OAuth is the real identity boundary.

Useful environment variables:

DEMO_WEB_ONLY=0                  # default: disable the browser-token gate
DEMO_REQUIRE_LOGIN=0             # default locally; set to 1 to require OAuth
DEMO_WEB_TOKEN_TTL_SECONDS=7200  # token lifetime
DEMO_WEB_GATE_SECRET=...         # optional stable signing secret
DEMO_USAGE_HASH_SECRET=...       # stable Space Secret for pseudonymous quota IDs
DEMO_DAILY_FREE_REQUESTS=10      # non-PRO daily generation limit
DEMO_DEFAULT_BACKEND=ggml        # ggml or torch
DEMO_AVAILABLE_BACKENDS=ggml,torch
DEMO_GGML_QUANT=BF16
USAGE_BUCKET_MOUNT_PATH=/data    # attached HF Storage Bucket mount path
USAGE_DB_FILENAME=faster-qwen3-tts-usage.sqlite3
USAGE_DB_PATH=/data/usage.sqlite3 # optional full path override

Set DEMO_USAGE_HASH_SECRET as a Space Secret before using the persistent bucket. If it is not set, the app falls back to DEMO_WEB_GATE_SECRET; if neither is set, quota IDs are still private but rotate on restart, resetting per-user counts.

To inspect usage from the bucket DB:

SELECT u.username, u.is_pro, d.day, d.count
FROM usage_daily d
LEFT JOIN usage_users u USING (user_key)
ORDER BY d.day DESC, d.count DESC;

Bucket setup:

hf buckets create HuggingFaceM4/faster-qwen3-tts-usage

For the hosted HuggingFaceM4 Space, attach the private HuggingFaceM4/faster-qwen3-tts-usage bucket in the Space settings under Storage Buckets with mount path /data and read-write access. The andito/faster-qwen3-tts-usage bucket is only needed for a personal andito-owned Space or fork.

You can also attach the org bucket with huggingface_hub:

from huggingface_hub import HfApi, Volume

api = HfApi()
api.set_space_volumes(
    "HuggingFaceM4/faster-qwen3-tts-demo",
    volumes=[
        Volume(
            type="bucket",
            source="HuggingFaceM4/faster-qwen3-tts-usage",
            mount_path="/data",
        ),
    ],
)
api.restart_space("HuggingFaceM4/faster-qwen3-tts-demo", factory_reboot=True)