Spaces:

dattazigzag
/

kokoro_test

Sleeping

App Files Files Community

kokoro_test / README.md

dattazigzag

Update README.md

c8a01ff verified 4 months ago

preview code

raw

history blame contribute delete

31.4 kB

	---
	license: apache-2.0
	title: kokoro_test
	sdk: gradio
	sdk_version: 5.25.2
	emoji: 🐨
	colorFrom: yellow
	colorTo: red
	pinned: true
	short_description: a gardio for kokoro
	thumbnail: >-
	https://cdn-uploads.huggingface.co/production/uploads/67ff88a5f2572583c60fd4cd/PkEkbfhPiKf8e_LQrfeuc.png
	---

	# README

	## Model Source

	## Kokoro-82M

	1. [Hugging Face](https://huggingface.co/hexgrad/Kokoro-82M)
	2. [Github](https://github.com/hexgrad/kokoro/tree/main?tab=Apache-2.0-1-ov-file)

	## Kokoro-onnx

	1. [Github](https://github.com/thewh1teagle/kokoro-onnx)

	## About Kokoro

	Kokoro ("heart" or "spirit" in Japanese) is an open-weight TTS model with only 82 million parameters. Despite its small size, it delivers impressive voice quality across multiple languages and voices.
	The model was trained exclusively on permissive/non-copyrighted audio data and IPA phoneme labels, making it suitable for both commercial and personal projects under the Apache 2.0 license.

	### Key Features (😊)

	1. __Lightweight Architecture__: Just 82M parameters, allowing for efficient inference
	2. __Multilingual Support__: 9 languages including English, Spanish, Japanese, and more
	3. __Multiple Voices__: 60+ voices across different languages and genders
	4. __Freedom to use commercially and privately__: kokooro - Apache 2.0 Licensed & kokoro-onnx: MIT
	5. __Strong Performance__: Competitive quality with much larger models. __Continuations__ and __annotations__ are great !! ✅
	6. __Near near real-time performance__: if models or voices are not changed, then loading time of optimized f32 version model is 2 secs . This time is not "audio generation time".
	7. Can control __speed__ of speaking and use __breaks__ as a feature for the underlying model to get sentence breakpoints and thus audio features per sentence.
	8. ‼️ No Voice Cloning ‼️ ---- 💬 _who wants that are you crazy? I'm okay with that_
	9. ‼️ No German (DE) at the moment ‼️ . But please [check below](#voices-summary) for a list of CURRENLY available languages.

	> kokoro-onnx specific

	1. ** _Even more faster performance and near real-time on macOS M sries chips_
	2. Can mix genders and voices (__Blending__) for interesting results

	```bash
	nicole: np.ndarray = kokoro.get_voice_style("af_nicole")
	michael: np.ndarray = kokoro.get_voice_style("am_michael")
	blend = np.add(nicole * (50 / 100), michael * (50 / 100))
	```

	[img tbd]

	3. Can get _runtime providers_ via built in API and use them during runtime to get even better performance

	```bash
	# To list providers, simply activte your venv
	>>> import onnxruntime
	>>> onnxruntime.get_all_providers()

	# you will see something like ...
	['TensorrtExecutionProvider', 'CUDAExecutionProvider', 'MIGraphXExecutionProvider', 'ROCMExecutionProvider', 'OpenVINOExecutionProvider', 'DnnlExecutionProvider', 'VitisAIExecutionProvider', 'QNNExecutionProvider', 'NnapiExecutionProvider', 'VSINPUExecutionProvider', 'JsExecutionProvider', 'CoreMLExecutionProvider', 'ArmNNExecutionProvider', 'ACLExecutionProvider', 'DmlExecutionProvider', 'RknpuExecutionProvider', 'WebNNExecutionProvider', 'WebGpuExecutionProvider', 'XnnpackExecutionProvider', 'CANNExecutionProvider', 'AzureExecutionProvider', 'CPUExecutionProvider']
	```

	So for mac M1 systems, we can either use that provider `'CoreMLExecutionProvider'` directly in the program in a session

	```bash
	session = onnxruntime.InferenceSession(
	model, providers=['CoreMLExecutionProvider', 'CPUExecutionProvider']
	)
	```

	or, use it during runtime

	```bash
	ONNX_PROVIDER="CoreMLExecutionProviderr" python main_kokoro_onnx.py
	```

	3. We get extra log levels. This gives you better visbilityu with what's going on under the hood and not just your program ...

	```bsah
	import logging
	...
	logging.getLogger(kokoro_onnx.__name__).setLevel("DEBUG")
	```

	4. Have lightweight options:
	1. [kokoro-v1.0.onnx](https://github.com/thewh1teagle/kokoro-onnx/releases/download/model-files-v1.0/kokoro-v1.0.onnx): (310MB): optimized f32 version
	2. [kokoro-v1.0.fp16.onnx](https://github.com/thewh1teagle/kokoro-onnx/releases/download/model-files-v1.0/kokoro-v1.0.fp16.onnx): (169MB): optimized f16 version
	3. [kokoro-v1.0.int8.onnx](https://github.com/thewh1teagle/kokoro-onnx/releases/download/model-files-v1.0/kokoro-v1.0.int8.onnx): (88MB): optimized int8 version
	5. You get specific GPU versions of onnyx models but for mac it always runs on GPU.

	```bsah
	pip install -U kokoro-onnx[gpu]
	# gpu version is sufficient only for Linux and Windows. macOS works with GPU by default
	```

	### In summary

	\| Feature \| Kokoro \| Kokoro-ONNX \|
	\|---------\|--------\|-------------\|
	\| Architecture \| PyTorch-based with HuggingFace integration \| ONNX-optimized for inference speed \|
	\| Model Loading \| `KPipeline(lang_code, repo_id, device)` downloads models from HuggingFace \| `Kokoro.from_session(session, bin_file)` using local ONNX and voice files \|
	\| Language Codes \| Short codes (`"a"` = en-us, `"b"` = en-gb) \| Standard language codes (`"en-us"`, `"en-gb"`) \|
	\| Audio Generation \| Generator pattern that yields `(graphemes, phonemes, audio)` chunks \| Single API call returning `(samples, sample_rate)` \|
	\| Text Processing \| Supports various split patterns (`r"\n+"`, `(?<=[.!?])\s+`) \| No built-in splitting (must be handled manually if needed) \|
	\| Hardware Acceleration \| Auto-detects CUDA/MPS/CPU \| Explicitly configure providers (CoreML, CUDA, CPU) via environment variables \|
	\| Phonemization \| Handles internally as part of generator pattern \| Separate `tokenizer.phonemize()` function (optional usage) \|
	\| Memory Management \| Streams audio in chunks, better for memory \| Generates entire audio at once (can be an issue for long texts) \|
	\| Voice Data \| Downloads voice models as needed \| Uses pre-bundled voice binary file \|
	\| Error Handling \| Detailed error handling for each generation stage \| Simpler error handling for the single API call \|
	\| Implementation Example \| `generator = pl(text, voice, speed, split_pattern); for i, (gs, ps, audio) in enumerate(generator): ...` \| `samples, sample_rate = kokoro.create(text, voice, speed, lang)` \|

	---

	## Kokoro's generated audio format

	1. __Sample Rate__: Fixed at 24kHz (24000 Hz)
	2. __Channels__: Mono (single channel)
	3. __Data Format__ (`Dtype Int8`, `Int16`, etc.): Selectable during saving data to a file. By default it uses 16-bit integer format (Int16)

	> In kokoo-onyx methods, the sample rate can be grabbed from the function and thus can be used for playback or file saving with prorper formatting where as in pure kokoro method, we use the known hard codded sample rate ...
	> In kokoo-onyx methods, the deafult data format is HQ 32bit floating pts.

	### Test System Info

	The model has been tested on the following system:

	> Datta's mac

	```txt
	OS: macOS 15.3.2 24D81 arm64
	CPU: Apple M3 Max
	Memory: 3371MiB / 36864MiB
	Python: 3.12
	```

	> Lower-end systems should also be capable of running the model effectively due to its lightweight architecture.

	### Repository flile structure

	```txt
	.
	├── LICENSE
	├── README.md
	├── assets/
	├── audio_exports/
	├── examples
	│ ├── 01_kk_play_save.py
	│ └── 02_kk_onnx_play_save.py
	├── extras
	│ ├── device_selection.py
	│ ├── get_sound_device_info.py
	│ ├── main_kokoro_intractive.py
	│ └── save_to_disk_and_then_play.py
	├── kokoro_gradio.py
	├── kokoro_gradio_client_example.py
	├── kokoro_onnx_basic_main.py
	├── kokoro_onnx_gradio.py
	├── kokoro_onnx_gradio_client_example.py
	├── onnx_deps
	│ ├── download_kokoro-onnx_deps.sh
	│ ├── kokoro-v1.0.fp16-gpu.onnx
	│ ├── kokoro-v1.0.fp16.onnx
	│ ├── kokoro-v1.0.int8.onnx
	│ ├── kokoro-v1.0.onnx
	│ └── voices-v1.0.bin
	├── pyproject.toml
	└── uv.lock
	```

	1. `assets/`: Assets for `README.md`
	2. `audio_exports/`: Dir where all the scripts export and save their audio from TTS, on disk
	3. `examples/`: Dir that has the two main headless python scripts for using pure kokoro (`01_kk_play_save.py`) or kokoro-onnx (`01_kk_play_save.py`). _**These scripts can be used as boiler plates or starting points for implementation_
	4. `extras/device_selection.py`: shows how to use kokoro runtime device (`CPU/CUDA/MLP/CPU`) - _not same for kokoro-onnx_
	5. `extras/get_sound_device_info.py`: shows how python `sounddevice` library can be used to idenify available sound card devices
	6. `extras/main_kokoro_intractive.py`: and cli interactive tool, using kokoro py lib, for testing language, voice and sentece combo for TTS. It was done befior eusing the g-radio versions but I left them there as it has a nice TUI, I like. But do not expect to do the same for kokoro-onnx (not the goal to make a pretty TUI tool).
	7. `extras/save_to_disk_and_then_play.py`: [TBD] Demo for showing how to sae TTS audio data to wav file, then load and do audio transformation, specifically to embedd playback soundcard features directly into the file. Might be helpful when multiple `soundfile` py lib can not be used for audio streaming and may need to play audio using a system levelk player like `afplay` (mac) or `aplay` (Linux), programatically from python ...
	8. `onnx_deps`: various onnx model files and voice pt files as .bin.
	9. `kokoro_gradio.py`: kokoro basic exmaple using gradio web gui as a playground.
	10. `kokoro_gradio_client_example.py`: example implementation to show how to interact with gradio kokoro server via API.
	11. `kokoro_onnx_gradio.py`: kokoro-onnx basic exmaple using gradio web gui as a playground.
	12. `kokoro_onnx_gradio_client_example.py`: example implementation to show how to interact with gradio kokoro-onnx server via API.

	---

	## Option 1: Install from scratch

	```bash
	# Make sure you have uv installed
	uv init -p 3.12

	# Create and activate virtual environment
	source .venv/bin/activate.fish # or use your shell-specific activation command

	# Install dependencies
	uv add kokoro kokoro-onnx misaki[ja] misaki[zh] soundfile sounddevice pip colorama numpy torch scipy gradio gradio-client

	# Extra
	uv add
	```

	## Option 1: Quick Install from project.toml

	```bash
	# Make sure your virtual environment is activated
	# source .venv/bin/activate.fish # or use your shell-specific activation command

	uv pip install -e .
	```

	## For kokoro-onnx, download the models

	```bash
	# ** For kokoro-onnx, donwload locally onnyx models and voices
	cd onnx_deps

	# INT8 (88 MB):
	wget https://github.com/thewh1teagle/kokoro-onnx/releases/download/model-files-v1.0/kokoro-v1.0.int8.onnx
	# FP16 (169 MB):
	wget https://github.com/thewh1teagle/kokoro-onnx/releases/download/model-files-v1.0/kokoro-v1.0.fp16.onnx
	# FP32 (326 MB):
	wget https://github.com/thewh1teagle/kokoro-onnx/releases/download/model-files-v1.0/kokoro-v1.0.onnx

	# ** For kokoro-onnx, donwload voices
	wget https://github.com/thewh1teagle/kokoro-onnx/releases/download/model-files-v1.0/voices-v1.0.bin
	```

	or run

	```bash
	cd onnx_deps
	./onnx_deps/download_kokoro-onnx_deps.sh

	# By default the model version is set to v1.0

	# Default version
	# VERSION="v1.0"

	# In the future you can just pass a version / tag to the script to download all .onnx and .bin for that specific release
	# e.g.: ./onnx_deps/download_kokoro-onnx_deps.sh v1.1
	```

	## Usage 1: Basic Usage for pure kokoro with gradio-gui

	Implementedin in gardio for playing around ...

	```bash
	# make sure you have venv activated !!
	python3 kokoro_gradio.py

	# or run on MacOS Apple Silicon GPU Acceleration
	PYTORCH_ENABLE_MPS_FALLBACK=1 python3 kokoro_gradio.py
	```

	> __Pure kokoro limitations__
	>
	> _Honestly these are not a bigie_
	>
	> 1. Doesn't include voice blending
	> 2. Programmatic assignment for GPU (process IO provider) not available
	> 3. Model DEBUG info not available

	![alt text](<assets/Screenshot 2025-04-14 at 19.13.43.png>)

	## Usage 1: Scripted bare metal usage for pure kokoro from hexgrid

	### Test 1: process, combine, play and then save

	1. Load language once and do not load again as it takes some time (approx. `1-2.5 secs`).
	2. Moreover the self induced constraint (assumption) here is that we won't switch language in between
	3. Check how long does it take to generate / process a chunk of audio; first being a single line sentence. Then play it from memory (from audio data buffer) and then save it to disk for reviewing ...
	4. Then process 1st multi-line text (with pargraphs in it). Then play it from memory (from audio data buffer) and then save it to disk for reviewing ...
	5. Then process the 2nd multi-line text (also with pargraphs in it). Then also play it from memory (from audio data buffer) and then save it to disk for reviewing ...
	6. Each tie tts is carried out (text is processed and audio data buffer is generated), the pipeline wasn't recreated. the pipeline for pure kokoro, only needs to be created when language needs to be changed. voice can be changed and for that pipeline doesn;t need to be initiated (whihc is tiny bot time consumeing)
	7. __Extra__ 😉: _I added audio transformation for stream playback and file saving. Meaning, you an match your sound card's samplerate, bitrate, can control gain, and specify outchannel/channels all from within the code, under the hood it uses [sounddevice.Stream API](https://python-sounddevice.readthedocs.io/en/0.5.1/api/streams.html#sounddevice.Stream)._

	```bash
	PYTORCH_ENABLE_MPS_FALLBACK=1 python3 examples/01_kk_play_save.py
	```

	### Result

	```bash
	# Pt. 1
	Loading pipeline ...
	Process took: 2.333237 secs.

	...

	# Pt. 3. (Single line with first time pipeline loaded)
	Processing Single Line Text
	Initializing generator Process took: 0.000012 secs.
	Chunk 1 creation took: 0.000024 secs.
	Streaming audio... (audio file length): 5.547685 secs.
	Saving file to disk took: 0.006516 secs.
	file size: 259 KB

	# Pt. 4 (Multi line sentences with reusing pipeline)
	Initializing generator Process took: 0.000006 secs
	Chunk 1 of 3 took: 0.000029 secs.
	Chunk 2 of 3 took: 0.000024 secs.
	Chunk 3 of 3 took: 0.000028 secs
	Combining took: 0.000361 secs.
	Streaming audio... (combined audio file length): 34.360400 secs.
	Saving file to disk took: 0.013127 secs.
	file size: 1.6 MB

	# Pt. 5 Multi line sentences with the previous context while reusing the 1st loaded pipeline)
	Initializing generator Process took: 0.000007 secs.
	Chunk 1 of 3 took: 0.000023 secs.
	Chunk 2 of 3 took: 0.000022 secs.
	Chunk 3 of 3 took: 0.000024 secs.
	Chunk 3 of 3 took: 0.000030 secs.
	Combining took: 0.000704 secs.
	Streaming audio... (combined audio file length): 43.408353 secs.
	Saving file to disk took: 0.015742 secs.
	file size: 2.1 MB
	```

	> Check Figma for more details

	---

	## Extra

	```bash
	cd extra
	PYTORCH_ENABLE_MPS_FALLBACK=1 python3 main_kokoro_intractive.py
	```

	### Interactive Commands

	When running the application, you can use these commands:

	1. `lang?` or `l?` - Display currently set language code
	2. `voice?` or `v?` - Display currently set voice
	3. `playback?` or `p?` - Display playback options
	4. `set lang to [code]` - Change language code to one from the list
	5. `set voice to [name]` - Change voice to one from the list
	6. `set playback to [mode]` - Change playback mode (`file` or `stream`)

	### Notes

	1. In playback mode `file`, it saves the generated audio data to a file and plays back on default soundcard / playback device
	2. In playback mode `stream`, it saves the generated audio data too but before that plays back the data from the memory through the default soundcard / playback device. TBH, difference is not at all significant.

	In file mode, in one example, total time from generating to file saving to the file file laoding tok 0.0032 sec

	```bash
	Current Settings:
	Language: h
	Voice: hf_alpha
	Speed: 1
	Playback: file


	...
	Initializing generator ...
	Process took: 0.0000 secs.
	Generator initialized with voice: hf_alpha
	Speed: 1
	[0] TEXT(graphemes): It was a Sprawl voice and a Sprawl joke. The Chatsubo was a bar for professional expatriates; you could drink there for a week and never hear two words in Japanese.

	(Phonemes): ɪt wʌzɐ spɹˈɔːl vˈYs and ɐ spɹˈɔːl ʤˈQk. ðə ʧatsjˈuːbQ wʌzɐ bˈɑː fɔː pɹəfˈɛʃənəl ɛkspˈatɹɪˌAts; juː kʊd dɹˈɪŋk ðeə fəɹə wˈiːk and nˈɛvə hˈiə tˈuː wˈɜːdz ɪn ʤˌapənˈiːz.

	Writing audio file (For immediate playback): audio_exports/0hf_alpha.wav
	Writing audio file (For immediate playback): audio_exports/0hf_alpha.wav
	Process took: 0.0032 secs.
	Success ✅
	Playing audio file: audio_exports/0hf_alpha.wav
	....
	```

	And in, stream mode, it took 0.0011 sec (incl. file saving in the background ...)

	```bash
	Initializing generator ...
	Process took: 0.0000 secs.
	Generator initialized with voice: hf_alpha
	Speed: 1
	[0] TEXT(graphemes): It was a Sprawl voice and a Sprawl joke. The Chatsubo was a bar for professional expatriates; you could drink there for a week and never hear two words in Japanese.

	(Phonemes): ɪt wʌzɐ spɹˈɔːl vˈYs and ɐ spɹˈɔːl ʤˈQk. ðə ʧatsjˈuːbQ wʌzɐ bˈɑː fɔː pɹəfˈɛʃənəl ɛkspˈatɹɪˌAts; juː kʊd dɹˈɪŋk ðeə fəɹə wˈiːk and nˈɛvə hˈiə tˈuː wˈɜːdz ɪn ʤˌapənˈiːz.

	Writing audio file (for recording purposes): audio_exports/0hf_alpha.wav
	Process took: 0.0017 secs.
	Success ✅
	Streaming audio...
	```

	## Usage 2: Basic Usage For pure kokoro-onnx with gradio-gui

	```bash
	ONNX_PROVIDER="CoreMLExecutionProvider" python3 kokoro_onnx_gradio.py
	```

	![alt text](<assets/Screenshot 2025-04-14 at 19.27.25.png>)

	## Usage 2: Scripted bare metal usage for kokoro-onnx

	```bash
	# On mac
	ONNX_PROVIDER="CoreMLExecutionProvider" python3 examples/02_kk_onnx_play_save.py
	```

	### Result

	```bash
	Model loading took: 3.503696 secs.
	ONNX model loaded with provider: CoreMLExecutionProvider
	Loading voice data from: /Users/saurabhdatta/Documents/Projects/VW/ArtificialAugmentation2025/tts_tests/kokoro_test/onnx_deps/voices-v1.0.bin
	Voice loading took: 0.002724 secs.


	Processing Single Line Text
	Generating audio for text:
	The sky above the port was the color of television, tuned to a dead channel.

	Audio generation took: 2.039143 secs.
	Playing audio on 2 channels at 48000Hz \| Duration: 4.33 seconds
	...
	Saving audio file...
	Saving took: 0.011106 secs.


	Processing Multi-Line Text 1
	Generating audio for text:
	Once upon a time, there was a little girl who lived in a village near the forest. Whenever she went out, the little girl wore a red riding cloak, so everyone in the village called her Little Red Riding Hood.

	One morning, Little Red Riding Hood asked her mother if she could go to visit her grandmother as it had been awhile since they'd seen each other.

	"That's a good idea," her mother said. "It's such a lovely day for a walk in the forest. Take this basket of fresh bread and butter to your grandmother, and remember - don't talk to strangers on the way!"

	Audio generation took: 5.621120 secs.
	Playing audio on 2 channels at 48000Hz \| Duration: 30.14 seconds
	...
	Saving audio file...
	Saving took: 0.030169 secs.


	Processing Multi-Line Text 2
	Generating audio for text:
	Little Red Riding Hood promised to be careful and set off immediately. The forest was dense and deep, with sunlight filtering through the leaves. Birds sang cheerfully as she walked along the path.

	Suddenly, she met a wolf. "Hello, little girl," said the wolf in a voice as sweet as honey. "Where are you going all alone in the woods?"

	Little Red Riding Hood didn't know that wolves could be dangerous, so she replied, "I'm going to visit my grandmother who lives on the other side of the forest."

	The wolf smiled wickedly. "What a coincidence! I was just heading that way myself. Why don't you take the long path with all the beautiful flowers? I'll take the short path and meet you there."


	Audio generation took: 6.558173 secs.
	Starting playback \| Duration: 37.95 seconds
	...
	Saving audio file...
	Saving took: 0.033002 secs.
	```

	---

	> 💡 For more detailed benchmark side by side, please checkout the [figma space ](https://www.figma.com/board/6tqMcW6uoGlxPVknI0biAO/Artificial-Augmentation---Ongoings?node-id=101-2815&t=vqCFuLhMcO6RZa5R-4)...


	# Voices Summary

	## Lang codes

	```txt
	# 🇺🇸 'a' => American English, 'b' => British English
	# 'e' => Spanish es
	# 'f' => French fr-fr
	# 'h' => Hindi hi
	# 'i' => Italian it
	# 'j' => Japanese: pip install misaki[ja]
	# 'p' => Brazilian Portuguese pt-br
	# 'z' => Mandarin Chinese: pip install misaki[zh]
	```

	\| Kokoro Code \| Standard Language Code \| Language Description \|
	\|-------------\|------------------------\|----------------------\|
	\| `a` \| `en-us` \| 🇺🇸 American English \|
	\| `b` \| `en-gb` \| 🇬🇧 British English \|
	\| `e` \| `es` \| 🇪🇸 Spanish (Spain) \|
	\| `f` \| `fr-fr` \| 🇫🇷 French (France) \|
	\| `h` \| `hi` \| 🇮🇳 Hindi \|
	\| `i` \| `it` \| 🇮🇹 Italian \|
	\| `j` \| `ja` \| 🇯🇵 Japanese \|
	\| `p` \| `pt-br` \| 🇧🇷 Brazilian Portuguese \|
	\| `z` \| `zh` \| 🇨🇳 Mandarin Chinese \|

	\| Name \| Traits \| Target Quality \| Training Duration \| Overall Grade \| SHA256 \|
	\| ---- \| ------ \| -------------- \| ----------------- \| ------------- \| ------ \|
	\| af\_heart \| 🚺❤️ \| \| \| A \| `0ab5709b` \|
	\| af_alloy \| 🚺 \| B \| MM minutes \| C \| `6d877149` \|
	\| af_aoede \| 🚺 \| B \| H hours \| C+ \| `c03bd1a4` \|
	\| af_bella \| 🚺🔥 \| A \| HH hours \| A- \| `8cb64e02` \|
	\| af_jessica \| 🚺 \| C \| MM minutes \| D \| `cdfdccb8` \|
	\| af_kore \| 🚺 \| B \| H hours \| C+ \| `8bfbc512` \|
	\| af_nicole \| 🚺🎧 \| B \| HH hours \| B- \| `c5561808` \|
	\| af_nova \| 🚺 \| B \| MM minutes \| C \| `e0233676` \|
	\| af_river \| 🚺 \| C \| MM minutes \| D \| `e149459b` \|
	\| af_sarah \| 🚺 \| B \| H hours \| C+ \| `49bd364e` \|
	\| af_sky \| 🚺 \| B \| _M minutes_ 🤏 \| C- \| `c799548a` \|
	\| am_adam \| 🚹 \| D \| H hours \| F+ \| `ced7e284` \|
	\| am_echo \| 🚹 \| C \| MM minutes \| D \| `8bcfdc85` \|
	\| am_eric \| 🚹 \| C \| MM minutes \| D \| `ada66f0e` \|
	\| am_fenrir \| 🚹 \| B \| H hours \| C+ \| `98e507ec` \|
	\| am_liam \| 🚹 \| C \| MM minutes \| D \| `c8255075` \|
	\| am_michael \| 🚹 \| B \| H hours \| C+ \| `9a443b79` \|
	\| am_onyx \| 🚹 \| C \| MM minutes \| D \| `e8452be1` \|
	\| am_puck \| 🚹 \| B \| H hours \| C+ \| `dd1d8973` \|
	\| am_santa \| 🚹 \| C \| _M minutes_ 🤏 \| D- \| `7f2f7582` \|

	<br>
	<details>
	<summary>More VOICE deatils (from hexgrad - maintainers of kokoro in HF)</summary>

	- 🇺🇸 [American English](#american-english): 11F 9M
	- 🇬🇧 [British English](#british-english): 4F 4M
	- 🇯🇵 [Japanese](#japanese): 4F 1M
	- 🇨🇳 [Mandarin Chinese](#mandarin-chinese): 4F 4M
	- 🇪🇸 [Spanish](#spanish): 1F 2M
	- 🇫🇷 [French](#french): 1F
	- 🇮🇳 [Hindi](#hindi): 2F 2M
	- 🇮🇹 [Italian](#italian): 1F 1M
	- 🇧🇷 [Brazilian Portuguese](#brazilian-portuguese): 1F 2M

	For each voice, the given grades are intended to be estimates of the quality and quantity of its associated training data, both of which impact overall inference quality.

	Subjectively, voices will sound better or worse to different people.

	Support for non-English languages may be absent or thin due to weak G2P and/or lack of training data. Some languages are only represented by a small handful or even just one voice (French).

	Most voices perform best on a "goldilocks range" of 100-200 tokens out of ~500 possible. Voices may perform worse at the extremes:
	- Weakness on short utterances, especially less than 10-20 tokens. Root cause could be lack of short-utterance training data and/or model architecture. One possible inference mitigation is to bundle shorter utterances together.
	- Rushing on long utterances, especially over 400 tokens. You can chunk down to shorter utterances or adjust the `speed` parameter to mitigate this.

	Target Quality
	- How high quality is the reference voice? This grade may be impacted by audio quality, artifacts, compression, & sample rate.
	- How well do the text labels match the audio? Text/audio misalignment (e.g. from hallucinations) will lower this grade.

	Training Duration
	- How much audio was seen during training? Smaller durations result in a lower overall grade.
	- 10 hours <= HH hours < 100 hours
	- 1 hour <= H hours < 10 hours
	- 10 minutes <= MM minutes < 100 minutes
	- 1 minute <= _M minutes_ 🤏 < 10 minutes

	### American English

	- `lang_code='a'` in [`misaki[en]`](https://github.com/hexgrad/misaki)
	- espeak-ng `en-us` fallback

	\| Name \| Traits \| Target Quality \| Training Duration \| Overall Grade \| SHA256 \|
	\| ---- \| ------ \| -------------- \| ----------------- \| ------------- \| ------ \|
	\| af\_heart \| 🚺❤️ \| \| \| A \| `0ab5709b` \|
	\| af_alloy \| 🚺 \| B \| MM minutes \| C \| `6d877149` \|
	\| af_aoede \| 🚺 \| B \| H hours \| C+ \| `c03bd1a4` \|
	\| af_bella \| 🚺🔥 \| A \| HH hours \| A- \| `8cb64e02` \|
	\| af_jessica \| 🚺 \| C \| MM minutes \| D \| `cdfdccb8` \|
	\| af_kore \| 🚺 \| B \| H hours \| C+ \| `8bfbc512` \|
	\| af_nicole \| 🚺🎧 \| B \| HH hours \| B- \| `c5561808` \|
	\| af_nova \| 🚺 \| B \| MM minutes \| C \| `e0233676` \|
	\| af_river \| 🚺 \| C \| MM minutes \| D \| `e149459b` \|
	\| af_sarah \| 🚺 \| B \| H hours \| C+ \| `49bd364e` \|
	\| af_sky \| 🚺 \| B \| _M minutes_ 🤏 \| C- \| `c799548a` \|
	\| am_adam \| 🚹 \| D \| H hours \| F+ \| `ced7e284` \|
	\| am_echo \| 🚹 \| C \| MM minutes \| D \| `8bcfdc85` \|
	\| am_eric \| 🚹 \| C \| MM minutes \| D \| `ada66f0e` \|
	\| am_fenrir \| 🚹 \| B \| H hours \| C+ \| `98e507ec` \|
	\| am_liam \| 🚹 \| C \| MM minutes \| D \| `c8255075` \|
	\| am_michael \| 🚹 \| B \| H hours \| C+ \| `9a443b79` \|
	\| am_onyx \| 🚹 \| C \| MM minutes \| D \| `e8452be1` \|
	\| am_puck \| 🚹 \| B \| H hours \| C+ \| `dd1d8973` \|
	\| am_santa \| 🚹 \| C \| _M minutes_ 🤏 \| D- \| `7f2f7582` \|

	### British English

	- `lang_code='b'` in [`misaki[en]`](https://github.com/hexgrad/misaki)
	- espeak-ng `en-gb` fallback

	\| Name \| Traits \| Target Quality \| Training Duration \| Overall Grade \| SHA256 \|
	\| ---- \| ------ \| -------------- \| ----------------- \| ------------- \| ------ \|
	\| bf_alice \| 🚺 \| C \| MM minutes \| D \| `d292651b` \|
	\| bf_emma \| 🚺 \| B \| HH hours \| B- \| `d0a423de` \|
	\| bf_isabella \| 🚺 \| B \| MM minutes \| C \| `cdd4c370` \|
	\| bf_lily \| 🚺 \| C \| MM minutes \| D \| `6e09c2e4` \|
	\| bm_daniel \| 🚹 \| C \| MM minutes \| D \| `fc3fce4e` \|
	\| bm_fable \| 🚹 \| B \| MM minutes \| C \| `d44935f3` \|
	\| bm_george \| 🚹 \| B \| MM minutes \| C \| `f1bc8122` \|
	\| bm_lewis \| 🚹 \| C \| H hours \| D+ \| `b5204750` \|

	### Japanese

	- `lang_code='j'` in [`misaki[ja]`](https://github.com/hexgrad/misaki)
	- Total Japanese training data: H hours

	\| Name \| Traits \| Target Quality \| Training Duration \| Overall Grade \| SHA256 \| CC BY \|
	\| ---- \| ------ \| -------------- \| ----------------- \| ------------- \| ------ \| ----- \|
	\| jf_alpha \| 🚺 \| B \| H hours \| C+ \| `1bf4c9dc` \| \|
	\| jf_gongitsune \| 🚺 \| B \| MM minutes \| C \| `1b171917` \| [gongitsune](https://github.com/koniwa/koniwa/blob/master/source/tnc/tnc__gongitsune.txt) \|
	\| jf_nezumi \| 🚺 \| B \| _M minutes_ 🤏 \| C- \| `d83f007a` \| [nezuminoyomeiri](https://github.com/koniwa/koniwa/blob/master/source/tnc/tnc__nezuminoyomeiri.txt) \|
	\| jf_tebukuro \| 🚺 \| B \| MM minutes \| C \| `0d691790` \| [tebukurowokaini](https://github.com/koniwa/koniwa/blob/master/source/tnc/tnc__tebukurowokaini.txt) \|
	\| jm_kumo \| 🚹 \| B \| _M minutes_ 🤏 \| C- \| `98340afd` \| [kumonoito](https://github.com/koniwa/koniwa/blob/master/source/tnc/tnc__kumonoito.txt) \|

	### Mandarin Chinese

	- `lang_code='z'` in [`misaki[zh]`](https://github.com/hexgrad/misaki)
	- Total Mandarin Chinese training data: H hours

	\| Name \| Traits \| Target Quality \| Training Duration \| Overall Grade \| SHA256 \|
	\| ---- \| ------ \| -------------- \| ----------------- \| ------------- \| ------ \|
	\| zf_xiaobei \| 🚺 \| C \| MM minutes \| D \| `9b76be63` \|
	\| zf_xiaoni \| 🚺 \| C \| MM minutes \| D \| `95b49f16` \|
	\| zf_xiaoxiao \| 🚺 \| C \| MM minutes \| D \| `cfaf6f2d` \|
	\| zf_xiaoyi \| 🚺 \| C \| MM minutes \| D \| `b5235dba` \|
	\| zm_yunjian \| 🚹 \| C \| MM minutes \| D \| `76cbf8ba` \|
	\| zm_yunxi \| 🚹 \| C \| MM minutes \| D \| `dbe6e1ce` \|
	\| zm_yunxia \| 🚹 \| C \| MM minutes \| D \| `bb2b03b0` \|
	\| zm_yunyang \| 🚹 \| C \| MM minutes \| D \| `5238ac22` \|

	### Spanish

	- `lang_code='e'` in [`misaki[en]`](https://github.com/hexgrad/misaki)
	- espeak-ng `es`

	\| Name \| Traits \| SHA256 \|
	\| ---- \| ------ \| ------ \|
	\| ef_dora \| 🚺 \| `d9d69b0f` \|
	\| em_alex \| 🚹 \| `5eac53f7` \|
	\| em_santa \| 🚹 \| `aa8620cb` \|

	### French

	- `lang_code='f'` in [`misaki[en]`](https://github.com/hexgrad/misaki)
	- espeak-ng `fr-fr`
	- Total French training data: <11 hours

	\| Name \| Traits \| Target Quality \| Training Duration \| Overall Grade \| SHA256 \| CC BY \|
	\| ---- \| ------ \| -------------- \| ----------------- \| ------------- \| ------ \| ----- \|
	\| ff_siwis \| 🚺 \| B \| <11 hours \| B- \| `8073bf2d` \| [SIWIS](https://datashare.ed.ac.uk/handle/10283/2353) \|

	### Hindi

	- `lang_code='h'` in [`misaki[en]`](https://github.com/hexgrad/misaki)
	- espeak-ng `hi`
	- Total Hindi training data: H hours

	\| Name \| Traits \| Target Quality \| Training Duration \| Overall Grade \| SHA256 \|
	\| ---- \| ------ \| -------------- \| ----------------- \| ------------- \| ------ \|
	\| hf_alpha \| 🚺 \| B \| MM minutes \| C \| `06906fe0` \|
	\| hf_beta \| 🚺 \| B \| MM minutes \| C \| `63c0a1a6` \|
	\| hm_omega \| 🚹 \| B \| MM minutes \| C \| `b55f02a8` \|
	\| hm_psi \| 🚹 \| B \| MM minutes \| C \| `2f0f055c` \|

	### Italian

	- `lang_code='i'` in [`misaki[en]`](https://github.com/hexgrad/misaki)
	- espeak-ng `it`
	- Total Italian training data: H hours

	\| Name \| Traits \| Target Quality \| Training Duration \| Overall Grade \| SHA256 \|
	\| ---- \| ------ \| -------------- \| ----------------- \| ------------- \| ------ \|
	\| if_sara \| 🚺 \| B \| MM minutes \| C \| `6c0b253b` \|
	\| im_nicola \| 🚹 \| B \| MM minutes \| C \| `234ed066` \|

	### Brazilian Portuguese

	- `lang_code='p'` in [`misaki[en]`](https://github.com/hexgrad/misaki)
	- espeak-ng `pt-br`

	\| Name \| Traits \| SHA256 \|
	\| ---- \| ------ \| ------ \|
	\| pf_dora \| 🚺 \| `07e4ff98` \|
	\| pm_alex \| 🚹 \| `cf0ba8c5` \|
	\| pm_santa \| 🚹 \| `d4210316` \|
	</details>

	----

	## Training

	> Why on earth you would wanna do that ... 🤔 ?

	\| Training Costs \| v0.19 \| v1.0 \| Total \|
	\| --- \| --- \| --- \| --- \|
	\| In A100 80GB GPU hours \| 500 \| 500 \| 1000 \|
	\| Average hourly rate \| $0.80/h \| $1.20/h \| $1/h \|
	\| In USD \| $400 \| $600 \| $1000 \|