# AI use cases for Paradee (sahilmahendrakar/paradee) Repo: [sahilmahendrakar/paradee](https://github.com/sahilmahendrakar/paradee) | Use cases: 4 | Page: https://dedupedai.com/use-cases/sahilmahendrakar-paradee/ Paradee is a small 7.88M-parameter English text-to-speech model, distilled from Kokoro-82M, that speaks one voice and runs on CPU. This file holds every use case of this repo. Each one starts with a "## " heading and ends with its URL. - Add offline English speech to a Python app: https://dedupedai.com/use-cases/sahilmahendrakar-paradee/add-offline-english-speech-to-python-app/ - Ship a browser voice with no WebGPU or download: https://dedupedai.com/use-cases/sahilmahendrakar-paradee/ship-browser-voice-without-webgpu/ - Call the Paradee ONNX graph from any runtime: https://dedupedai.com/use-cases/sahilmahendrakar-paradee/call-paradee-onnx-graph-from-any-runtime/ - Study how a TTS model was distilled: https://dedupedai.com/use-cases/sahilmahendrakar-paradee/study-distillation-of-a-tts-model/ ## How to add offline English speech to a Python app with Paradee Repo: [sahilmahendrakar/paradee](https://github.com/sahilmahendrakar/paradee) | Difficulty: Intermediate | Time: 15 minutes | Basis: From the README Install Paradee from GitHub and generate a WAV file from text on a CPU. **Who it is for:** A Python developer building a script, tool or prototype that needs spoken English output without a cloud TTS account or a GPU. **The problem:** Cloud speech APIs need keys and network calls, and larger open TTS models are heavy to run on a laptop. **The outcome:** A WAV file of a sentence, generated locally by the Paradee model in a few seconds per sentence. ### Steps 1. Run `pip install git+https://github.com/sahilmahendrakar/paradee` to install the package. 2. Run `python -m paradee "Paradee is a small voice that runs anywhere." -o hello.wav` to generate a test file. 3. In your Python code, import Paradee and SAMPLE_RATE from Paradee, create `tts = Paradee()`, and call `tts("your text")` to get the audio array. (inferred, check the README) 4. Write the array to a file with soundfile using `sf.write("out.wav", audio, SAMPLE_RATE)`. (inferred, check the README) ### Prompt to paste into your agent ```text Install the GitHub repo sahilmahendrakar/paradee with pip, then generate speech for the text and save it to . Expect a WAV file at 24 kHz that plays back as the spoken sentence. Verify the file exists and is longer than zero seconds before reporting success. ``` ### Caveat Paradee speaks only one voice, Kokoro's af_heart, and the GitHub description's 18x speed figure is for version 1.0 while the README gives 62x for version 1.1. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/sahilmahendrakar-paradee/add-offline-english-speech-to-python-app/ ## How to ship a browser voice with no WebGPU or download with Paradee Repo: [sahilmahendrakar/paradee](https://github.com/sahilmahendrakar/paradee) | Difficulty: Advanced | Time: 1 hour | Basis: From the README Use Paradee's browser phonemizer and ONNX file to speak text inside a web page. **Who it is for:** A front-end developer building a Chrome extension or web app that needs speech but cannot require WebGPU or a large model download. **The problem:** Most in-browser TTS needs WebGPU or a big download, which rules out many users and pages. **The outcome:** A browser feature that turns text into speech using the light Paradee model, following the same setup as the Chickadee extension. ### Steps 1. Clone the repo with `git clone https://github.com/sahilmahendrakar/paradee` and open the `web/` folder. (inferred, check the README) 2. Use `web/misaki.js` to convert text to misaki-style phonemes, because browser phonemizers such as kokoro-js write some sounds differently. 3. Load the ONNX file from the Hugging Face model page into a browser ONNX runtime and pass phoneme ids as `input_ids` with `speed` set to 1.0. (inferred, check the README) 4. Read the waveform output, which is float32 at 24 kHz, and play it with the Web Audio API. ### Prompt to paste into your agent ```text Clone sahilmahendrakar/paradee, read web/misaki.js and the ONNX input spec in the README, then write a minimal browser page that speaks the text . Expect audio playback at 24 kHz with no WebGPU. Verify that the phoneme ids start and end with pad token 0 and total at most 512 before running inference. ``` ### Caveat The README does not give a browser runtime or loading code, so the ONNX loading step is an inference and the author's Chickadee extension is the only documented browser use. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/sahilmahendrakar-paradee/ship-browser-voice-without-webgpu/ ## How to call the Paradee ONNX graph from any runtime with Paradee Repo: [sahilmahendrakar/paradee](https://github.com/sahilmahendrakar/paradee) | Difficulty: Advanced | Time: 45 minutes | Basis: From the README Feed phoneme ids and speed into the ONNX graph and get 24 kHz audio back. **Who it is for:** A backend or ML engineer who wants to serve TTS from a non-Python stack or a CPU server, using the raw model file. **The problem:** Wrapping the Python package is awkward in some services, so a direct ONNX interface is needed. **The outcome:** A service that takes phoneme ids and returns a float32 waveform at 24 kHz from the single ONNX graph. ### Steps 1. Download the int8 ONNX model from huggingface.co/sahilmahendrakar/Paradee-8M-v1.1. 2. Get the phoneme vocabulary from `config.json` on the Hub, and build `input_ids` as int64 `[1, T]` with a 0 pad token at each end, at most 512 tokens. 3. Create `speed` as float32 `[1]`, where 1.0 is normal speed. 4. Run the graph and read the `waveform` output, float32 `[1, samples]` at 24 kHz. ### Prompt to paste into your agent ```text Load the Paradee 1.1 int8 ONNX file from sahilmahendrakar/Paradee-8M-v1.1 with an ONNX runtime, build input_ids from phonemes in config.json with 0 pads, and run inference on the text . Expect a float32 waveform at 24 kHz. Verify the output shape is [1, samples] and that the sample count is greater than zero. ``` ### Caveat The phonemes must come from misaki or be converted to misaki's spelling, since the model was trained only on misaki output and browser phonemizers can cut word accuracy sharply. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/sahilmahendrakar-paradee/call-paradee-onnx-graph-from-any-runtime/ ## How to study how a TTS model was distilled with Paradee Repo: [sahilmahendrakar/paradee](https://github.com/sahilmahendrakar/paradee) | Difficulty: Advanced | Time: 2 hours | Basis: From the README Read the paper and training code to reproduce a teacher-student speech model on one laptop. **Who it is for:** A researcher or student who wants to learn how a large TTS model was compressed into a small one. **The problem:** Distillation steps for speech models are rarely laid out end to end with code that runs on a single machine. **The outcome:** A working understanding of the corpus, text-side training, decoder training and assembly steps, plus a local run of one stage. ### Steps 1. Read the paper at arxiv.org/abs/2610.06817 and the How it works section of the README. 2. Open the `training/` folder and the `training/data/` sentences, which come from WikiText-103. 3. Review how the text side, decoder and assembly steps are split, noting that the decoder is trained with spectrogram losses and then adversarial training. 4. Run one stage from `training/` on your laptop, since the README says the code runs on one laptop. ### Prompt to paste into your agent ```text Clone sahilmahendrakar/paradee, read the How it works section and the training/ folder, then summarize the four distillation steps and the parameter split for the text side and decoder. Expect a short summary with the 4.23M and 3.65M parameter counts. Verify that every number you report appears in the README or paper before finalizing. ``` ### Caveat The paper describes version 1.0 and the 1.1 decoder head is only summarized in the README, so the code for 1.1 may differ from the paper's method. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/sahilmahendrakar-paradee/study-distillation-of-a-tts-model/