Teaching Kokoro to speak nerd
Ready.
Text is processed on the server, not saved. No sensitive information. 3 requests/minute per visitor; shared quotas apply.
Vocabulary
- CUDA
- coodda
- PostgreSQL
- Postgress
- Postgres
- Postgress
- PyTorch
- Py Torch
- bfloat16
- bee float sixteen
- BF16
- bee float sixteen
- WAV
- wave
- API
- eigh pee eye
- GPU
- gee pee you
- CPU
- see pee you
- DGX
- dee gee ex
- SSH
- ess ess aitch
- HTTP
- aitch tee tee pee
- HTTPS
- aitch tee tee pee ess
- TTS
- tee tee ess
- STT
- ess tee tee
- LLM
- el el em
- ML
- em el
- CLI
- see el eye
- SDK
- ess dee kay
- KV
- kay vee
- SQL
- sequel
- JSON
- jason
- YAML
- yamul
- TOML
- tomul
- SQLite
- sequel light
- tmux
- tee mux
- Tailscale
- tail scale
- Neovim
- neo vim
- JEPA
- jeppa
- VJEPA
- vee jeppa
- TiDAR
- tee dar
- JAX
- jacks
- XLA
- ex el eigh
- DeepSeek
- deep seek
- Qwen
- kwen
- SGLang
- ess gee lang
- vLLM
- vee el el em
- LoRA
- lora
- QLoRA
- cue lora
- MoE
- mow
- DINO
- dino
- ViT
- vit
- MAE
- em eigh ee
- BERT
- bert
- CodeBERT
- code bert
- MiniCPM
- mini see pee em
- RoPE
- rope
- NCCL
- nickel
- cuDNN
- coo dee en en
- cuBLAS
- coo blass
- cuFFT
- coo eff eff tee
- TensorRT
- tensor ar tee
- Triton
- triton
- NVCC
- en vee see see
- PTX
- pee tee ex
- SASS
- sass
- ARM
- arm
- ARM64
- arm sixty four
- FP16
- eff pee sixteen
- FP32
- eff pee thirty two
- FP8
- eff pee eight
- NVFP4
- en vee eff pee four
- E4M3
- ee four em three
- E5M2
- ee five em two
- SDPA
- ess dee pee eigh
- CFG
- see eff gee
- GQA
- gee cue eigh
- MQA
- em cue eigh
- RMSNorm
- ar em ess norm
- SwiGLU
- swig loo
- ReLU
- ree loo
- GELU
- gelu
- SiLU
- see loo
- FlashAttention
- flash attention
- AdamW
- adam double you
- GGUF
- gee gee you eff
- GGML
- gee gee em el
- ONNX
- onyx
- NumPy
- num pie
- SciPy
- sigh pie
- Matplotlib
- mat plot lib
- TensorFlow
- tensor flow
- HuggingFace
- hugging face
- safetensors
- safe tensors
- Nsight
- en sight
- TMA
- tee em eigh
- WGMMA
- double you gee em em eigh
- UMMA
- you em em eigh
- CUTLASS
- cutlass
- cuTENSOR
- coo tensor
- cuSPARSELt
- coo sparse el tee
- cuSOLVER
- coo solver
- AI
- eigh eye
- GPT
- gee pee tee
- OpenAI
- open eigh eye
- OpenRouter
- open router
- NVIDIA
- envidia
- GitHub
- git hub
- GitLab
- git lab
- GPUs
- gee pee yous
- CPUs
- see pee yous
- GB
- gigabytes
- GiB
- gibby bytes
- MB
- megabytes
- MiB
- mebibytes
- SSD
- ess ess dee
- pee dee eff
- PNG
- pee en gee
- JPEG
- jay peg
- UI
- you eye
- URL
- you ar el
- ID
- eye dee
- S3
- ess three
- B300
- bee three hundred
- GB10
- gee bee ten
- K3
- kay three
- macOS
- mac oh ess
- SIGReg
- sig reg
- LeJEPA
- luh jeppa
- RunPod
- run pod
- UnderLeaf
- under leaf
- VM
- vee em
- DNS
- dee en ess
- CI
- see eye
- JSONL
- jason el
- CodeSearchNet
- code search net
- GIANT
- giant
- SUPER
- super
- ICLR
- eye see el ar
- NeurIPS
- new rips
- VibeVoice
- vibe voice
- SigLIP
- sig lip
133 terms · Download JSON
The trick: lie to the phonemizer
- Technical terms: inconsistent default pronunciation.
- Fix: dictionary of spoken spellings, applied before phonemization.
- No fine-tuning. No new weights. No LLM at inference.
Text“CUDA uses an API”
Dictionaryspoken spellings
Misaki G2Ptext → phonemes
Kokorophonemes → audio
WAV24 kHz
What is Kokoro?
- Kokoro-82M: 82M parameters; open weights; Apache-2.0.
- Architecture: StyleTTS 2 + ISTFTNet decoder.
- English phonemizer: Misaki, with Espeak fallback.
- This demo: American English;
af_heartvoice. - Local GPU; fixed public gateway; one synthesis slot; no queue.
Build your own list
- Ask your agent to scan your conversation JSON files or SQLite database. Count technical terms; rank by frequency. That's how I built my seed list.
- Have it write every term's intended pronunciation as phonemes, using your model's phoneme alphabet.
- Encode for your TTS frontend. This demo uses pronounceable spellings that produce those phonemes, not raw IPA.
- Test the phonemizer output. Listen. Fix exceptions. Exclude personal names and secrets before sharing.