| --- |
| license: apache-2.0 |
| library_name: ggml |
| pipeline_tag: text-to-speech |
| language: |
| - en |
| base_model: owensong/Inflect-Nano-v1 |
| base_model_relation: quantized |
| tags: |
| - text-to-speech |
| - gguf |
| - quantized |
| - inflect-nano |
| - english |
| --- |
| |
| <p align="center"> |
| <img src="https://huggingface.co/owensong/Inflect-Nano-v1/resolve/main/assets/inflect-nano-banner.png" alt="Inflect-Nano banner" width="100%"> |
| </p> |
|
|
| # Inflect-Nano GGUF Quantized Weights |
|
|
| This repository contains quantized GGUF weights of [Inflect-Nano-v1 model](https://huggingface.co/owensong/Inflect-Nano-v1) for the [InflectNanoTTS C++/GGML runtime](https://github.com/remixer-dec/InflectNanoTTS.cpp). |
| The original model was published by its author, [owensong](https://huggingface.co/owensong) under Apache-2.0 license. Inflect-Nano-v1 aims to deliver ultra-lightweight english text to speech to edge devices. |
| |
| The goal of this port was to run TTS on an ESP32 device and it has been [achieved](https://github.com/remixer-dec/PocketInkOS/tree/master/src/tts/). From a cold start to a spoken single word, the full neural vocoder takes around 26 seconds. In Griffin-Lim mode it takes around 6 seconds and is good enough for short phrases. |
|
|
| [](https://t.me/genaimon) |
|
|
| ## Repository Contents |
|
|
| ```text |
| acoustic/ |
| inflect_acoustic_f16.gguf |
| inflect_acoustic_q2_k.gguf |
| inflect_acoustic_q3_k.gguf |
| inflect_acoustic_q3_k_dynamic.gguf |
| inflect_acoustic_q4_0.gguf |
| inflect_acoustic_q4_0_E.gguf |
| inflect_acoustic_q4_k.gguf |
| inflect_acoustic_q5_0.gguf |
| inflect_acoustic_q5_k.gguf |
| inflect_acoustic_q6_k.gguf |
| inflect_acoustic_q8_0.gguf |
| vocoder/ |
| inflect_vocoder_f16.gguf |
| inflect_vocoder_q4_0.gguf |
| inflect_vocoder_q4_0_E.gguf |
| inflect_vocoder_q5_k.gguf |
| inflect_vocoder_q8_0.gguf |
| cmudict.bin |
| cmudict.idx |
| ``` |
|
|
| Quality drop becomes significant lower than q3_k, others work relatively well. |
| |
| The `inflect_acoustic_q4_0_E.gguf` and `inflect_vocoder_q4_0_E.gguf` files are experimental low-memory quants. They keep biases in fp16 and the last layer is also quantized to q4_0, making the pair small enough to fit in an ESP32's 8MB PSRAM. |
|
|
| ## Usage |
|
|
| Build the runtime from the [InflectNanoTTS.cpp](https://github.com/remixer-dec/InflectNanoTTS.cpp) project: |
|
|
| ```bash |
| ./tools/build.sh |
| ``` |
|
|
| Run inference: |
|
|
| ```bash |
| build/linux-arm64/inflect-nano \ |
| -a acoustic/inflect_acoustic_q4_0.gguf \ |
| -v vocoder/inflect_vocoder_q4_0.gguf \ |
| -d cmudict.bin \ |
| -t "Hello, this is a test." \ |
| -o output.wav |
| ``` |
|
|
| Adjust the binary path for your platform, for example `build/macos-arm64/inflect-nano` on Apple Silicon. |
|
|
| **cmudict.bin** is a compact binary pronunciation dictionary used by the text frontend. |
|
|
| **cmudict.idx** is an index file for `cmudict.bin` that makes dictionary initialization faster. |
|
|
| It is derived from CMU Pronouncing Dictionary: English words mapped to ARPAbet phoneme sequences. The project converts that text dictionary into a faster binary lookup format so the runtime can avoid parsing the full text dictionary at startup. |
|
|
| Its purpose is to turn English words into phoneme IDs before acoustic inference. In practice: input text -> normalize/tokenize -> dictionary lookup for pronunciations -> phoneme/tone/lang IDs -> acoustic model -> mel -> vocoder audio. |
|
|
| cmudict.bin was obtained by converting [cmudict.rep from the original repo](https://huggingface.co/owensong/Inflect-Nano-v1/blob/main/third_party/tiny_tts_frontend/tiny_tts/text/cmudict.rep) with [this script](https://github.com/remixer-dec/InflectNanoTTS.cpp/blob/master/tools/compile_cmudict.py): |
|
|
| ```python |
| python tools/compile_cmudict.py \ |
| third_party/tiny_tts_frontend/tiny_tts/text/cmudict.rep \ |
| cmudict.bin |
| ``` |
|
|
