HangryNugget | Hangry Labs, Game Servers, AI Systems

Fast local speech

KokoroTTS

A compact 82M-parameter TTS engine with 54 voices, nine language prefixes, a browser UI, and a practical HTTP API.

Choose it when: Fast narration, long-form generation, first-time local TTS users, and lightweight integrations.

Compute: Runs on CPU; NVIDIA GPU acceleration is optional.

Memory fit8 GB is comfortably in range; the lower floor still needs a published lab run.

Speed / RTFNo verified 5060 Ti / 5070 Ti RTF is published yet.

Evidence:Compatibility guidance ↗

82M parametersCPU friendly54 voicesUI + API

First command

docker run -p 7860:7860 hangrylabs/kokorotts:v0.2

Repository ↗ Examples ↗

Stable multilingual TTS

MeloTTS

A dependable multilingual speech engine packaged with full and English-focused images, UI, API, and sentence streaming.

Choose it when: English, Spanish, French, Chinese, Japanese, or Korean speech with predictable controls and integration routes.

Compute: CPU and NVIDIA GPU modes are supported.

Memory fit8 GB is the conservative picker tier; a measured minimum is not published yet.

Speed / RTFNo verified 5060 Ti / 5070 Ti RTF is published yet.

Evidence:Compatibility guidance ↗

6 language familiesStreamingUI + APICPU fallback

First command

docker run -p 8888:8888 --gpus all hangrylabs/melotts:latest

Repository ↗ Examples ↗

Massively multilingual voice lab

OmniVoiceTTS

A broad TTS system for 600+ languages with voice design, reference cloning, reusable profiles, and OpenAI-compatible speech routes.

Choose it when: Multilingual products, expressive voice design, consented voice cloning, and OpenAI-compatible local speech.

Compute: GPU recommended; CPU fallback needs roughly 2–7 GB RAM depending on cloning and ASR use.

Memory fit5.31 GiB peak reserved: 6 GB practical floor, 8 GB recommended.

Speed / RTFRTF 0.24–0.35 on RTX 5060 Ti; 0.26–0.33 on RTX 5070 Ti (about 2.8–4.2× real time).

Evidence:Hangry Labs measured ↗

6 GB minimum600+ languagesVoice cloningOpenAI-compatible API

First command

docker run -p 7861:7861 --gpus all hangrylabs/omnivoicetts:latest

Repository ↗ Examples ↗

Creative and true-to-life speech

VoxCPMTTS

VoxCPM2 packaged for local generation, natural voice design, reference cloning, and transcript-guided cloning.

Choose it when: Creative voices, realistic consented cloning, multilingual output, and advanced local speech experiments.

Compute: A large model: CPU fallback exists; a capable NVIDIA GPU is the practical path.

Memory fitAbout 8 GB according to the upstream VoxCPM2 reference.

Speed / RTFUpstream RTF ≈0.30 on RTX 4090; the minimum real-time lab GPU is not verified yet.

Evidence:Upstream reference ↗

~8 GB VRAMVoice designVoice cloningUI + API

First command

docker run -p 8808:8808 --gpus all hangrylabs/voxcpmtts:v0.1

Repository ↗

Speech to text

Qwen3-ASR-STT

A local ASR service for multilingual transcription, language detection, timestamps, microphone streaming, and OpenAI-compatible clients.

Choose it when: Transcription, captions, meeting audio, voice interfaces, and applications already using OpenAI-shaped APIs.

Compute: NVIDIA GPU workflow; the default 0.6B model is the lighter starting point and the 1.7B model trades memory for capability.

Memory fitMeasured startup: 4,620 MiB (0.6B) and 7,060 MiB (1.7B). Choose 6 GB or 8 GB respectively.

Speed / RTFTranscription speed depends on audio length, concurrency and backend; see the repository benchmark matrix.

Evidence:Hangry Labs measured ↗

4.62–6.89 GiB measuredMultilingual ASRRealtime sessionsOpenAI-compatible API

First command

docker run -p 8000:8000 --gpus all hangrylabs/qwen3-asr-stt:latest

Repository ↗

llama.cpp operations

llama.nodrama

A focused diagnostics interface for llama.cpp slots, KV cache behavior, and the operational details hidden behind a local LLM endpoint.

Choose it when: Understanding a running llama.cpp server, diagnosing capacity, and keeping local inference boring in the best way.

Compute: Hardware-neutral companion UI; requirements are determined by the llama.cpp server it observes.

Memory fitNo additional GPU memory requirement beyond the llama.cpp server it observes.

Speed / RTFA diagnostics interface, so audio real-time factor does not apply.

Evidence:Project architecture ↗

Slot diagnosticsKV cache insightLocal endpointNo-drama operations

First command

git clone https://github.com/hangry-labs/llama.nodrama.git

Repository ↗

Choose your language

Tools that work for you.
On hardware you control.

Browse the current projects

KokoroTTS

MeloTTS

OmniVoiceTTS

VoxCPMTTS

Qwen3-ASR-STT

llama.nodrama

KokoroTTS

MeloTTS

OmniVoiceTTS

VoxCPMTTS

Qwen3-ASR-STT

llama.nodrama

What are you trying to do?

Choose a category

What will run it? optional

How much GPU memory? optional

Choose a category to begin

KokoroTTS

MeloTTS

OmniVoiceTTS

VoxCPMTTS

Qwen3-ASR-STT

llama.nodrama

KokoroTTS

MeloTTS

OmniVoiceTTS

VoxCPMTTS

Qwen3-ASR-STT

llama.nodrama

Let’s get Docker ready for your computer

Which operating system do you use?

Which graphics hardware do you have? “Not sure” is fine

Tools that work for you.On hardware you control.

Browse the current projects

KokoroTTS

MeloTTS

OmniVoiceTTS

VoxCPMTTS

Qwen3-ASR-STT

llama.nodrama

KokoroTTS

MeloTTS

OmniVoiceTTS

VoxCPMTTS

Qwen3-ASR-STT

llama.nodrama

What are you trying to do?

Choose a category

What will run it? optional

How much GPU memory? optional

Choose a category to begin

KokoroTTS

MeloTTS

OmniVoiceTTS

VoxCPMTTS

Qwen3-ASR-STT

llama.nodrama

KokoroTTS

MeloTTS

OmniVoiceTTS

VoxCPMTTS

Qwen3-ASR-STT

llama.nodrama

Let’s get Docker ready for your computer

Which operating system do you use?

Which graphics hardware do you have? “Not sure” is fine

Tools that work for you.
On hardware you control.