woolball-xyz/woolball-server

Your own browser-based inference infrastructure by turning idle browsers into compute nodes.

114 commits

updated Feb 26, 2026

open.woolball.xyz

See the code

README

🧶 Woolball Server

Transform idle browsers into a powerful distributed AI inference network

Your own browser-based inference infrastructure by turning idle browsers into compute nodes

🚀 Quick Start • 📖 API Reference • 🛠️ Development • 💬 Discord

✨ What is Woolball?

Woolball Server is an open-source network server that orchestrates AI inference jobs across a distributed network of browser-based compute nodes. Instead of relying on expensive cloud infrastructure, harness the collective power of idle browsers to run AI models efficiently and cost-effectively.

🔗 Client side: Available in woolball-client
📋 Roadmap: Check our next steps

🎯 Supported AI Tasks

🔧 Provider	🎯 Task	🤖 Models	📊 Status
Transformers.js	🎤 Speech-to-Text	ONNX Models	✅ Ready
Transformers.js	🔊 Text-to-Speech	ONNX Models	✅ Ready
Kokoro.js	🔊 Text-to-Speech	ONNX Models	✅ Ready
Transformers.js	🌐 Translation	ONNX Models	✅ Ready
Transformers.js	📝 Text Generation	ONNX Models	✅ Ready
WebLLM	📝 Text Generation	MLC Models	✅ Ready
MediaPipe	📝 Text Generation	LiteRT Models	✅ Ready

🚀 Quick Start

Get up and running in under 2 minutes:

1️⃣ Clone & Deploy

git clone --branch deploy --single-branch --depth 1 https://github.com/woolball-xyz/woolball-server.git
cd woolball-server && docker compose up -d

2️⃣ Verify Setup

Open http://localhost:9000 to ensure at least one client node is connected.

3️⃣ Start Using the API

curl -X POST http://localhost:9002/api/v1/text-generation \
  -F 'input=[{"role":"user","content":"Hello! Can you explain what Woolball is?"}]' \
  -F "model=https://woolball.sfo3.cdn.digitaloceanspaces.com/gemma3-1b-it-int4-web.task" \
  -F "provider=mediapipe" \
  -F "maxTokens=200"

☁️ One-Click Deploy to DigitalOcean

Deploy Woolball to DigitalOcean App Platform with a single click:

🔧 What gets deployed:

🌐 Woolball Client: Frontend interface accessible via your app URL
🔌 Core API: RESTful API for AI inference jobs (/api route)
🔗 WebSocket Server: Real-time communication with browser nodes (/ws route)
⚙️ Background Service: Job orchestration and node management
📊 Redis Database: Managed Redis instance for caching and queues

🚀 After Deployment:

Your app will be available at https://your-app-name.ondigitalocean.app
API endpoint: https://your-app-name.ondigitalocean.app/api/v1
WebSocket: wss://your-app-name.ondigitalocean.app/ws

📖 API Reference

📖 Text Generation

Generate text with powerful language models

🤗 Transformers.js Provider

🤖 Available Models

Model	Quantization	Description
`HuggingFaceTB/SmolLM2-135M-Instruct`	`fp16`	Compact model for basic text generation
`HuggingFaceTB/SmolLM2-360M-Instruct`	`q4`	Balanced performance and size
`Mozilla/Qwen2.5-0.5B-Instruct`	`q4`	Efficient model for general tasks
`onnx-community/Qwen2.5-Coder-0.5B-Instruct`	`q8`	Specialized for code generation

💡 Example Usage

curl -X POST http://localhost:9002/api/v1/text-generation \
  -F 'input=[{"role":"system","content":"You are a helpful assistant."},{"role":"user","content":"What is the capital of Brazil?"}]' \
  -F "model=HuggingFaceTB/SmolLM2-135M-Instruct" \
  -F "dtype=fp16" \
  -F "max_new_tokens=250" \
  -F "temperature=0.7" \
  -F "do_sample=true"

⚙️ Parameters

Parameter	Type	Default	Description
`model`	string	-	🤖 Model ID (e.g., "HuggingFaceTB/SmolLM2-135M-Instruct")
`dtype`	string	-	🔧 Quantization level (e.g., "fp16", "q4")
`max_length`	number	20	📏 Maximum length the generated tokens can have (includes input prompt)
`max_new_tokens`	number	null	🆕 Maximum number of tokens to generate, ignoring prompt length
`min_length`	number	0	📐 Minimum length of the sequence to be generated (includes input prompt)
`min_new_tokens`	number	null	🔢 Minimum numbers of tokens to generate, ignoring prompt length
`do_sample`	boolean	false	🎲 Whether to use sampling; use greedy decoding otherwise
`num_beams`	number	1	🔍 Number of beams for beam search. 1 means no beam search
`temperature`	number	1.0	🌡️ Value used to modulate the next token probabilities
`top_k`	number	50	🔝 Number of highest probability vocabulary tokens to keep for top-k-filtering
`top_p`	number	1.0	📊 If < 1, only tokens with probabilities adding up to top_p or higher are kept
`repetition_penalty`	number	1.0	🔄 Parameter for repetition penalty. 1.0 means no penalty
`no_repeat_ngram_size`	number	0	🚫 If > 0, all ngrams of that size can only occur once

🤖 WebLLM Provider

🤖 Available Models

Model	Description
`DeepSeek-R1-Distill-Qwen-7B-q4f16_1-MLC`	DeepSeek R1 distilled model with reasoning capabilities
`DeepSeek-R1-Distill-Llama-8B-q4f16_1-MLC`	DeepSeek R1 distilled Llama-based model
`SmolLM2-1.7B-Instruct-q4f32_1-MLC`	Compact instruction-following model
`Llama-3.1-8B-Instruct-q4f32_1-MLC`	Meta's Llama 3.1 8B instruction model
`Qwen3-8B-q4f32_1-MLC`	Alibaba's Qwen3 8B model

💡 Example Usage

curl -X POST http://localhost:9002/api/v1/text-generation \
  -F 'input=[{"role":"system","content":"You are a helpful assistant."},{"role":"user","content":"What is the capital of Brazil?"}]' \
  -F "model=DeepSeek-R1-Distill-Qwen-7B-q4f16_1-MLC" \
  -F "provider=webllm" \
  -F "temperature=0.7" \
  -F "top_p=0.95"

⚙️ Parameters

Parameter	Type	Description
`model`	string	🤖 Model ID from MLC (e.g., "DeepSeek-R1-Distill-Qwen-7B-q4f16_1-MLC")
`provider`	string	🔧 Must be set to "webllm" when using WebLLM models
`context_window_size`	number	🪟 Size of the context window for the model
`sliding_window_size`	number	🔄 Size of the sliding window for attention
`attention_sink_size`	number	🎯 Size of the attention sink
`repetition_penalty`	number	🔄 Penalty for repeating tokens
`frequency_penalty`	number	📊 Penalty for token frequency
`presence_penalty`	number	👁️ Penalty for token presence
`top_p`	number	📈 If < 1, only tokens with probabilities adding up to top_p or higher are kept
`temperature`	number	🌡️ Value used to modulate the next token probabilities
`bos_token_id`	number	🏁 Beginning of sequence token ID (optional)

📱 MediaPipe Provider

🤖 Available Models

Model	Device Type	Description
`https://woolball.sfo3.cdn.digitaloceanspaces.com/gemma2-2b-it-cpu-int8.task`	CPU	Gemma2 2B model optimized for CPU inference
`https://woolball.sfo3.cdn.digitaloceanspaces.com/gemma2-2b-it-gpu-int8.bin`	GPU	Gemma2 2B model optimized for GPU inference
`https://woolball.sfo3.cdn.digitaloceanspaces.com/gemma3-1b-it-int4-web.task`	CPU/GPU	Gemma3 1B model with INT4 quantization
`https://woolball.sfo3.cdn.digitaloceanspaces.com/gemma3-4b-it-int4-web.task`	Web	Gemma3 4B model optimized for web deployment

💡 Example Usage

curl -X POST http://localhost:9002/api/v1/text-generation \
  -F 'input=[{"role":"system","content":"You are a helpful assistant."},{"role":"user","content":"Explain quantum computing in simple terms."}]' \
  -F "model=https://woolball.sfo3.cdn.digitaloceanspaces.com/gemma3-1b-it-int4-web.task" \
  -F "provider=mediapipe" \
  -F "maxTokens=500" \
  -F "temperature=0.7" \
  -F "topK=40" \
  -F "randomSeed=12345"

⚙️ Parameters

Parameter	Type	Description
`model`	string	🤖 Model ID for MediaPipe LiteRT models on DigitalOcean Spaces
`provider`	string	🔧 Must be set to "mediapipe" when using MediaPipe models
`maxTokens`	number	🔢 Maximum number of tokens to generate
`randomSeed`	number	🎲 Random seed for reproducible results
`topK`	number	🔝 Number of highest probability vocabulary tokens to keep for top-k-filtering
`temperature`	number	🌡️ Value used to modulate the next token probabilities

🎤 Speech Recognition

Convert audio to text with Whisper models

🤖 Available Models

Model	Quantization	Description
`onnx-community/whisper-large-v3-turbo_timestamped`	`q4`	🎯 High accuracy with timestamps
`onnx-community/whisper-small`	`q4`	⚡ Fast processing

💡 Example Usage

# 📁 Local file
curl -X POST http://localhost:9002/api/v1/speech-recognition \
  -F "input=@/path/to/your/file.mp3" \
  -F "model=onnx-community/whisper-large-v3-turbo_timestamped" \
  -F "dtype=q4" \
  -F "language=en" \
  -F "return_timestamps=true" \
  -F "stream=false"

# 🔗 URL
curl -X POST http://localhost:9002/api/v1/speech-recognition \
  -F "input=https://example.com/audio.mp3" \
  -F "model=onnx-community/whisper-large-v3-turbo_timestamped" \
  -F "dtype=q4" \
  -F "language=en" \
  -F "return_timestamps=true" \
  -F "stream=false"

# 📊 Base64
curl -X POST http://localhost:9002/api/v1/speech-recognition \
  -F "input=data:audio/mp3;base64,YOUR_BASE64_ENCODED_AUDIO" \
  -F "model=onnx-community/whisper-large-v3-turbo_timestamped" \
  -F "dtype=q4" \
  -F "language=en" \
  -F "return_timestamps=true" \
  -F "stream=false"

⚙️ Parameters

Parameter	Type	Description
`model`	string	🤖 Model ID from Hugging Face (e.g., "onnx-community/whisper-large-v3-turbo_timestamped")
`dtype`	string	🔧 Quantization level (e.g., "q4")
`return_timestamps`	boolean \| 'word'	⏰ Return timestamps ("word" for word-level). Default is `false`.
`stream`	boolean	📡 Stream results in real-time. Default is `false`.
`chunk_length_s`	number	📏 Length of audio chunks to process in seconds. Default is `0` (no chunking).
`stride_length_s`	number	🔄 Length of overlap between consecutive audio chunks in seconds. If not provided, defaults to `chunk_length_s / 6`.
`force_full_sequences`	boolean	🎯 Whether to force outputting full sequences or not. Default is `false`.
`language`	string	🌍 Source language (auto-detect if null). Use this to potentially improve performance if the source language is known.
`task`	null \| 'transcribe' \| 'translate'	🎯 The task to perform. Default is `null`, meaning it should be auto-detected.
`num_frames`	number	🎬 The number of frames in the input audio.

🔊 Text-to-Speech

Generate natural speech from text

🤗 Transformers.js (MMS Models)

🤖 Available Models

Language	Model	Flag
English	`Xenova/mms-tts-eng`	🇺🇸
Spanish	`Xenova/mms-tts-spa`	🇪🇸
French	`Xenova/mms-tts-fra`	🇫🇷
German	`Xenova/mms-tts-deu`	🇩🇪
Portuguese	`Xenova/mms-tts-por`	🇵🇹
Russian	`Xenova/mms-tts-rus`	🇷🇺
Arabic	`Xenova/mms-tts-ara`	🇸🇦
Korean	`Xenova/mms-tts-kor`	🇰🇷

💡 Example Usage

# Standard request
curl -X POST http://localhost:9002/api/v1/text-to-speech \
  -F "input=Hello, this is a test for text to speech." \
  -F "model=Xenova/mms-tts-eng" \
  -F "dtype=q8" \
  -F "stream=false"

# Streaming request
curl -X POST http://localhost:9002/api/v1/text-to-speech \
  -F "input=Hello, this is a test for streaming text to speech." \
  -F "model=Xenova/mms-tts-eng" \
  -F "dtype=q8" \
  -F "stream=true"

⚙️ Parameters

Parameter	Type	Description	Required For
`model`	string	🤖 Model ID	All providers
`dtype`	string	🔧 Quantization level (e.g., "q8")	All providers
`stream`	boolean	📡 Whether to stream the audio response. Default is `false`.	All providers

🐱 Kokoro (Premium Voices)

🤖 Available Models

Model	Quantization	Description
`onnx-community/Kokoro-82M-ONNX`	`q8`	High-quality English TTS with multiple voices
`onnx-community/Kokoro-82M-v1.0-ONNX`	`q8`	Alternative Kokoro model version

💡 Example Usage

# Standard request
curl -X POST http://localhost:9002/api/v1/text-to-speech \
  -F "input=Hello, this is a test using Kokoro voices." \
  -F "model=onnx-community/Kokoro-82M-ONNX" \
  -F "voice=af_nova" \
  -F "dtype=q8" \
  -F "stream=false"

# Streaming request
curl -X POST http://localhost:9002/api/v1/text-to-speech \
  -F "input=Hello, this is a test using Kokoro voices with streaming." \
  -F "model=onnx-community/Kokoro-82M-ONNX" \
  -F "voice=af_nova" \
  -F "dtype=q8" \
  -F "stream=true"

⚙️ Parameters

Parameter	Type	Description	Required For
`model`	string	🤖 Model ID	Required
`dtype`	string	🔧 Quantization level (e.g., "q8")	Required
`voice`	string	🎭 Voice ID (see below)	Required
`stream`	boolean	📡 Whether to stream the audio response. Default is `false`.	Optional

🎭 Available Voice Options

🇺🇸 American Voices

👩 Female: af_heart, af_alloy, af_aoede, af_bella, af_jessica, af_nova, af_sarah
👨 Male: am_adam, am_echo, am_eric, am_liam, am_michael, am_onyx

🇬🇧 British Voices

👩 Female: bf_emma, bf_isabella, bf_alice, bf_lily
👨 Male: bm_george, bm_lewis, bm_daniel, bm_fable

🌐 Translation

Translate between 200+ languages

🤖 Available Models

Model	Quantization	Description
`Xenova/nllb-200-distilled-600M`	`q8`	🌍 Multilingual translation model supporting 200+ languages

💡 Example Usage

curl -X POST http://localhost:9002/api/v1/translation \
  -F "input=Hello, how are you today?" \
  -F "model=Xenova/nllb-200-distilled-600M" \
  -F "dtype=q8" \
  -F "srcLang=eng_Latn" \
  -F "tgtLang=por_Latn"

🌍 Language Support

Uses FLORES200 format - supports 200+ languages!

⚙️ Parameters

Parameter	Type	Description
`model`	string	🤖 Model ID (e.g., "Xenova/nllb-200-distilled-600M")
`dtype`	string	🔧 Quantization level (e.g., "q8")
`srcLang`	string	🌍 Source language code in FLORES200 format (e.g., "eng_Latn")
`tgtLang`	string	🌍 Target language code in FLORES200 format (e.g., "por_Latn")

🛠️ Local Development

🐳 Docker Setup (Recommended)

git clone https://github.com/woolball-xyz/woolball-server.git
cd woolball-server && docker compose up --build -d

🌐 Service Endpoints

🔧 Service	🚪 Port	🔗 URL
🔌 WebSocket	9003	localhost:9003
🌐 API Server	9002	localhost:9002
👥 Client Demo	9000	localhost:9000

🔄 Network Flow

Network Architecture

🤝 Contributing

We welcome contributions! Here's how you can help:

🐛 Report bugs via GitHub Issues
💡 Suggest features in our Discord
🔧 Submit PRs for improvements
📖 Improve documentation

📄 License

This project is licensed under the AGPL-3.0 License - see the LICENSE file for details.

Made with ❤️ by the Woolball team

🌟 Star us on GitHub • 💬 Join Discord

wasm

web-ai

webgpu

Similar repos

woolball-xyz/woolball-client

Turn any open tab into an AI compute node

same org

runpod-labs/wandler

transformers.js inference server

sauravpanda/BrowserAI

Run local LLMs like llama, deepseek-distill, kokoro and more inside your browser

1,451

blajanclaudiu/fun

QuentinFuxa/WhisperLiveKit

Real-time, local speech-to-text with streaming ASR, speaker diarization, translation, and…

11,113

lenML/Speech-AI-Forge

🍦 Speech-AI-Forge is a project developed around TTS generation model, implementing an API Server…

1,420

travisvn/chatterbox-tts-api

Local, OpenAI-compatible text-to-speech (TTS) API using Chatterbox, enabling users to generate…

684

jordimas/ctranslate2-web-server

ctranslate2-web-server with OpenAI compatibility

Languages

100.0%

woolball-xyz/woolball-server

Your own browser-based inference infrastructure by turning idle browsers into compute nodes.

114 commits

updated Feb 26, 2026

open.woolball.xyz

See the code

README

🧶 Woolball Server

Transform idle browsers into a powerful distributed AI inference network

Your own browser-based inference infrastructure by turning idle browsers into compute nodes

🚀 Quick Start • 📖 API Reference • 🛠️ Development • 💬 Discord

✨ What is Woolball?

🔗 Client side: Available in woolball-client
📋 Roadmap: Check our next steps

🎯 Supported AI Tasks

🔧 Provider	🎯 Task	🤖 Models	📊 Status
Transformers.js	🎤 Speech-to-Text	ONNX Models	✅ Ready
Transformers.js	🔊 Text-to-Speech	ONNX Models	✅ Ready
Kokoro.js	🔊 Text-to-Speech	ONNX Models	✅ Ready
Transformers.js	🌐 Translation	ONNX Models	✅ Ready
Transformers.js	📝 Text Generation	ONNX Models	✅ Ready
WebLLM	📝 Text Generation	MLC Models	✅ Ready
MediaPipe	📝 Text Generation	LiteRT Models	✅ Ready

🚀 Quick Start

Get up and running in under 2 minutes:

1️⃣ Clone & Deploy

git clone --branch deploy --single-branch --depth 1 https://github.com/woolball-xyz/woolball-server.git
cd woolball-server && docker compose up -d

2️⃣ Verify Setup

Open http://localhost:9000 to ensure at least one client node is connected.

3️⃣ Start Using the API

curl -X POST http://localhost:9002/api/v1/text-generation \
  -F 'input=[{"role":"user","content":"Hello! Can you explain what Woolball is?"}]' \
  -F "model=https://woolball.sfo3.cdn.digitaloceanspaces.com/gemma3-1b-it-int4-web.task" \
  -F "provider=mediapipe" \
  -F "maxTokens=200"

☁️ One-Click Deploy to DigitalOcean

Deploy Woolball to DigitalOcean App Platform with a single click:

🔧 What gets deployed:

🌐 Woolball Client: Frontend interface accessible via your app URL
🔌 Core API: RESTful API for AI inference jobs (/api route)
🔗 WebSocket Server: Real-time communication with browser nodes (/ws route)
⚙️ Background Service: Job orchestration and node management
📊 Redis Database: Managed Redis instance for caching and queues

🚀 After Deployment:

Your app will be available at https://your-app-name.ondigitalocean.app
API endpoint: https://your-app-name.ondigitalocean.app/api/v1
WebSocket: wss://your-app-name.ondigitalocean.app/ws

📖 API Reference

📖 Text Generation

Generate text with powerful language models

🤗 Transformers.js Provider

🤖 Available Models

Model	Quantization	Description
`HuggingFaceTB/SmolLM2-135M-Instruct`	`fp16`	Compact model for basic text generation
`HuggingFaceTB/SmolLM2-360M-Instruct`	`q4`	Balanced performance and size
`Mozilla/Qwen2.5-0.5B-Instruct`	`q4`	Efficient model for general tasks
`onnx-community/Qwen2.5-Coder-0.5B-Instruct`	`q8`	Specialized for code generation

💡 Example Usage

curl -X POST http://localhost:9002/api/v1/text-generation \
  -F 'input=[{"role":"system","content":"You are a helpful assistant."},{"role":"user","content":"What is the capital of Brazil?"}]' \
  -F "model=HuggingFaceTB/SmolLM2-135M-Instruct" \
  -F "dtype=fp16" \
  -F "max_new_tokens=250" \
  -F "temperature=0.7" \
  -F "do_sample=true"

⚙️ Parameters

Parameter	Type	Default	Description
`model`	string	-	🤖 Model ID (e.g., "HuggingFaceTB/SmolLM2-135M-Instruct")
`dtype`	string	-	🔧 Quantization level (e.g., "fp16", "q4")
`max_length`	number	20	📏 Maximum length the generated tokens can have (includes input prompt)
`max_new_tokens`	number	null	🆕 Maximum number of tokens to generate, ignoring prompt length
`min_length`	number	0	📐 Minimum length of the sequence to be generated (includes input prompt)
`min_new_tokens`	number	null	🔢 Minimum numbers of tokens to generate, ignoring prompt length
`do_sample`	boolean	false	🎲 Whether to use sampling; use greedy decoding otherwise
`num_beams`	number	1	🔍 Number of beams for beam search. 1 means no beam search
`temperature`	number	1.0	🌡️ Value used to modulate the next token probabilities
`top_k`	number	50	🔝 Number of highest probability vocabulary tokens to keep for top-k-filtering
`top_p`	number	1.0	📊 If < 1, only tokens with probabilities adding up to top_p or higher are kept
`repetition_penalty`	number	1.0	🔄 Parameter for repetition penalty. 1.0 means no penalty
`no_repeat_ngram_size`	number	0	🚫 If > 0, all ngrams of that size can only occur once

🤖 WebLLM Provider

🤖 Available Models

Model	Description
`DeepSeek-R1-Distill-Qwen-7B-q4f16_1-MLC`	DeepSeek R1 distilled model with reasoning capabilities
`DeepSeek-R1-Distill-Llama-8B-q4f16_1-MLC`	DeepSeek R1 distilled Llama-based model
`SmolLM2-1.7B-Instruct-q4f32_1-MLC`	Compact instruction-following model
`Llama-3.1-8B-Instruct-q4f32_1-MLC`	Meta's Llama 3.1 8B instruction model
`Qwen3-8B-q4f32_1-MLC`	Alibaba's Qwen3 8B model

💡 Example Usage

curl -X POST http://localhost:9002/api/v1/text-generation \
  -F 'input=[{"role":"system","content":"You are a helpful assistant."},{"role":"user","content":"What is the capital of Brazil?"}]' \
  -F "model=DeepSeek-R1-Distill-Qwen-7B-q4f16_1-MLC" \
  -F "provider=webllm" \
  -F "temperature=0.7" \
  -F "top_p=0.95"

⚙️ Parameters

Parameter	Type	Description
`model`	string	🤖 Model ID from MLC (e.g., "DeepSeek-R1-Distill-Qwen-7B-q4f16_1-MLC")
`provider`	string	🔧 Must be set to "webllm" when using WebLLM models
`context_window_size`	number	🪟 Size of the context window for the model
`sliding_window_size`	number	🔄 Size of the sliding window for attention
`attention_sink_size`	number	🎯 Size of the attention sink
`repetition_penalty`	number	🔄 Penalty for repeating tokens
`frequency_penalty`	number	📊 Penalty for token frequency
`presence_penalty`	number	👁️ Penalty for token presence
`top_p`	number	📈 If < 1, only tokens with probabilities adding up to top_p or higher are kept
`temperature`	number	🌡️ Value used to modulate the next token probabilities
`bos_token_id`	number	🏁 Beginning of sequence token ID (optional)

📱 MediaPipe Provider

🤖 Available Models

Model	Device Type	Description
`https://woolball.sfo3.cdn.digitaloceanspaces.com/gemma2-2b-it-cpu-int8.task`	CPU	Gemma2 2B model optimized for CPU inference
`https://woolball.sfo3.cdn.digitaloceanspaces.com/gemma2-2b-it-gpu-int8.bin`	GPU	Gemma2 2B model optimized for GPU inference
`https://woolball.sfo3.cdn.digitaloceanspaces.com/gemma3-1b-it-int4-web.task`	CPU/GPU	Gemma3 1B model with INT4 quantization
`https://woolball.sfo3.cdn.digitaloceanspaces.com/gemma3-4b-it-int4-web.task`	Web	Gemma3 4B model optimized for web deployment

💡 Example Usage

curl -X POST http://localhost:9002/api/v1/text-generation \
  -F 'input=[{"role":"system","content":"You are a helpful assistant."},{"role":"user","content":"Explain quantum computing in simple terms."}]' \
  -F "model=https://woolball.sfo3.cdn.digitaloceanspaces.com/gemma3-1b-it-int4-web.task" \
  -F "provider=mediapipe" \
  -F "maxTokens=500" \
  -F "temperature=0.7" \
  -F "topK=40" \
  -F "randomSeed=12345"

⚙️ Parameters

Parameter	Type	Description
`model`	string	🤖 Model ID for MediaPipe LiteRT models on DigitalOcean Spaces
`provider`	string	🔧 Must be set to "mediapipe" when using MediaPipe models
`maxTokens`	number	🔢 Maximum number of tokens to generate
`randomSeed`	number	🎲 Random seed for reproducible results
`topK`	number	🔝 Number of highest probability vocabulary tokens to keep for top-k-filtering
`temperature`	number	🌡️ Value used to modulate the next token probabilities

🎤 Speech Recognition

Convert audio to text with Whisper models

🤖 Available Models

Model	Quantization	Description
`onnx-community/whisper-large-v3-turbo_timestamped`	`q4`	🎯 High accuracy with timestamps
`onnx-community/whisper-small`	`q4`	⚡ Fast processing

💡 Example Usage

# 📁 Local file
curl -X POST http://localhost:9002/api/v1/speech-recognition \
  -F "input=@/path/to/your/file.mp3" \
  -F "model=onnx-community/whisper-large-v3-turbo_timestamped" \
  -F "dtype=q4" \
  -F "language=en" \
  -F "return_timestamps=true" \
  -F "stream=false"

# 🔗 URL
curl -X POST http://localhost:9002/api/v1/speech-recognition \
  -F "input=https://example.com/audio.mp3" \
  -F "model=onnx-community/whisper-large-v3-turbo_timestamped" \
  -F "dtype=q4" \
  -F "language=en" \
  -F "return_timestamps=true" \
  -F "stream=false"

# 📊 Base64
curl -X POST http://localhost:9002/api/v1/speech-recognition \
  -F "input=data:audio/mp3;base64,YOUR_BASE64_ENCODED_AUDIO" \
  -F "model=onnx-community/whisper-large-v3-turbo_timestamped" \
  -F "dtype=q4" \
  -F "language=en" \
  -F "return_timestamps=true" \
  -F "stream=false"

⚙️ Parameters

Parameter	Type	Description
`model`	string	🤖 Model ID from Hugging Face (e.g., "onnx-community/whisper-large-v3-turbo_timestamped")
`dtype`	string	🔧 Quantization level (e.g., "q4")
`return_timestamps`	boolean \| 'word'	⏰ Return timestamps ("word" for word-level). Default is `false`.
`stream`	boolean	📡 Stream results in real-time. Default is `false`.
`chunk_length_s`	number	📏 Length of audio chunks to process in seconds. Default is `0` (no chunking).
`stride_length_s`	number	🔄 Length of overlap between consecutive audio chunks in seconds. If not provided, defaults to `chunk_length_s / 6`.
`force_full_sequences`	boolean	🎯 Whether to force outputting full sequences or not. Default is `false`.
`language`	string	🌍 Source language (auto-detect if null). Use this to potentially improve performance if the source language is known.
`task`	null \| 'transcribe' \| 'translate'	🎯 The task to perform. Default is `null`, meaning it should be auto-detected.
`num_frames`	number	🎬 The number of frames in the input audio.

🔊 Text-to-Speech

Generate natural speech from text

🤗 Transformers.js (MMS Models)

🤖 Available Models

Language	Model	Flag
English	`Xenova/mms-tts-eng`	🇺🇸
Spanish	`Xenova/mms-tts-spa`	🇪🇸
French	`Xenova/mms-tts-fra`	🇫🇷
German	`Xenova/mms-tts-deu`	🇩🇪
Portuguese	`Xenova/mms-tts-por`	🇵🇹
Russian	`Xenova/mms-tts-rus`	🇷🇺
Arabic	`Xenova/mms-tts-ara`	🇸🇦
Korean	`Xenova/mms-tts-kor`	🇰🇷

💡 Example Usage

# Standard request
curl -X POST http://localhost:9002/api/v1/text-to-speech \
  -F "input=Hello, this is a test for text to speech." \
  -F "model=Xenova/mms-tts-eng" \
  -F "dtype=q8" \
  -F "stream=false"

# Streaming request
curl -X POST http://localhost:9002/api/v1/text-to-speech \
  -F "input=Hello, this is a test for streaming text to speech." \
  -F "model=Xenova/mms-tts-eng" \
  -F "dtype=q8" \
  -F "stream=true"

⚙️ Parameters

Parameter	Type	Description	Required For
`model`	string	🤖 Model ID	All providers
`dtype`	string	🔧 Quantization level (e.g., "q8")	All providers
`stream`	boolean	📡 Whether to stream the audio response. Default is `false`.	All providers

🐱 Kokoro (Premium Voices)

🤖 Available Models

Model	Quantization	Description
`onnx-community/Kokoro-82M-ONNX`	`q8`	High-quality English TTS with multiple voices
`onnx-community/Kokoro-82M-v1.0-ONNX`	`q8`	Alternative Kokoro model version

💡 Example Usage

# Standard request
curl -X POST http://localhost:9002/api/v1/text-to-speech \
  -F "input=Hello, this is a test using Kokoro voices." \
  -F "model=onnx-community/Kokoro-82M-ONNX" \
  -F "voice=af_nova" \
  -F "dtype=q8" \
  -F "stream=false"

# Streaming request
curl -X POST http://localhost:9002/api/v1/text-to-speech \
  -F "input=Hello, this is a test using Kokoro voices with streaming." \
  -F "model=onnx-community/Kokoro-82M-ONNX" \
  -F "voice=af_nova" \
  -F "dtype=q8" \
  -F "stream=true"

⚙️ Parameters

Parameter	Type	Description	Required For
`model`	string	🤖 Model ID	Required
`dtype`	string	🔧 Quantization level (e.g., "q8")	Required
`voice`	string	🎭 Voice ID (see below)	Required
`stream`	boolean	📡 Whether to stream the audio response. Default is `false`.	Optional

🎭 Available Voice Options

🇺🇸 American Voices

👩 Female: af_heart, af_alloy, af_aoede, af_bella, af_jessica, af_nova, af_sarah
👨 Male: am_adam, am_echo, am_eric, am_liam, am_michael, am_onyx

🇬🇧 British Voices

👩 Female: bf_emma, bf_isabella, bf_alice, bf_lily
👨 Male: bm_george, bm_lewis, bm_daniel, bm_fable

🌐 Translation

Translate between 200+ languages

🤖 Available Models

Model	Quantization	Description
`Xenova/nllb-200-distilled-600M`	`q8`	🌍 Multilingual translation model supporting 200+ languages

💡 Example Usage

curl -X POST http://localhost:9002/api/v1/translation \
  -F "input=Hello, how are you today?" \
  -F "model=Xenova/nllb-200-distilled-600M" \
  -F "dtype=q8" \
  -F "srcLang=eng_Latn" \
  -F "tgtLang=por_Latn"

🌍 Language Support

Uses FLORES200 format - supports 200+ languages!

⚙️ Parameters

Parameter	Type	Description
`model`	string	🤖 Model ID (e.g., "Xenova/nllb-200-distilled-600M")
`dtype`	string	🔧 Quantization level (e.g., "q8")
`srcLang`	string	🌍 Source language code in FLORES200 format (e.g., "eng_Latn")
`tgtLang`	string	🌍 Target language code in FLORES200 format (e.g., "por_Latn")

🛠️ Local Development

🐳 Docker Setup (Recommended)

git clone https://github.com/woolball-xyz/woolball-server.git
cd woolball-server && docker compose up --build -d

🌐 Service Endpoints

🔧 Service	🚪 Port	🔗 URL
🔌 WebSocket	9003	localhost:9003
🌐 API Server	9002	localhost:9002
👥 Client Demo	9000	localhost:9000

🔄 Network Flow

Network Architecture

🤝 Contributing

We welcome contributions! Here's how you can help:

🐛 Report bugs via GitHub Issues
💡 Suggest features in our Discord
🔧 Submit PRs for improvements
📖 Improve documentation

📄 License

This project is licensed under the AGPL-3.0 License - see the LICENSE file for details.

Made with ❤️ by the Woolball team

🌟 Star us on GitHub • 💬 Join Discord

wasm

web-ai

webgpu

Similar repos

woolball-xyz/woolball-client

Turn any open tab into an AI compute node

same org

runpod-labs/wandler

transformers.js inference server

sauravpanda/BrowserAI

Run local LLMs like llama, deepseek-distill, kokoro and more inside your browser

1,451

blajanclaudiu/fun

QuentinFuxa/WhisperLiveKit

Real-time, local speech-to-text with streaming ASR, speaker diarization, translation, and…

11,113

lenML/Speech-AI-Forge

🍦 Speech-AI-Forge is a project developed around TTS generation model, implementing an API Server…

1,420

travisvn/chatterbox-tts-api

Local, OpenAI-compatible text-to-speech (TTS) API using Chatterbox, enabling users to generate…

684

jordimas/ctranslate2-web-server

ctranslate2-web-server with OpenAI compatibility

Languages

100.0%