LingTTS: Offline Chinese Neural Text-to-Speech for ESP32-S3
็ฎไฝไธญๆ่ฏดๆ | English
The model and engine are distributed through the community SDK on GitHub; this repository contains the model card, not standalone weights. It is not a Transformers-ready checkpoint.
LingTTS is a neural text-to-speech model for offline Mandarin speech synthesis on ESP32-S3-class devices. It generates speech locally from Chinese text, including stories, short passages and device information. The ESP-IDF SDK provides an integration interface.
Try the online demo โ the demo runs the same C code and model compiled for a PC server; use a board to evaluate device performance.
Can ESP32-S3 run neural TTS offline?
Yes. LingTTS runs the synthesis locally on ESP32-S3 without a cloud connection. The current community preview is 0.2.0-rc1. It supports Mandarin Chinese; English support is limited to letters and around 80 common words, not full English sentences.
ESP32-S3 performance and memory
The measurements below are for ESP32-S3 N16R8 at 240 MHz with ESP-IDF 5.5.6. Results depend on the application and memory configuration. RTF is synthesis time divided by generated audio duration; lower is faster.
| Metric | Measured result |
|---|---|
| Audio output | 16 kHz, 16-bit mono PCM, delivered in chunks |
| RTF | 0.42 with PSRAM / 0.48 using internal RAM, hello example including text front end; 0.53โ0.54 in the Wi-Fi demo |
| First PCM | 0.27โ0.32 seconds from invocation, not speaker playback latency |
| Memory | 2.63 MB PSRAM plus about 1 KB internal RAM, or 151 KB internal RAM without PSRAM; task stack, DMA and application memory are extra |
| Flash | About 2.6 MB for library, model and dictionary; fits within 4 MB Flash subject to application and partition needs |
| Model file | About 1.83 MB for one base voice; the optional second voice model is also about 1.83 MB |
The deployed model is quantized and packaged; a verified parameter count is not currently published. File size should not be read as parameter count.
See the performance details and integration guide (Chinese).
Try it on a board
The demo requires an ESP32-S3 N16R8 board, MAX98357A amplifier and speaker. Connect BCLK to GPIO 15, LRC to GPIO 16, and DIN to GPIO 7. Download the community demo package from Releases, extract it and run:
python -m pip install esptool==4.12.0
python flash.py --port COM5 # Linux / WSL: python3 flash.py --port /dev/ttyUSB0
Connect a phone to the LingTTS-Demo Wi-Fi hotspot (password lingtts-demo), open http://192.168.4.1, enter Chinese text and play it. For ESP-IDF integration, download the community SDK package from Releases and see the Chinese README.
Limitations
The current release offers a Mandarin voice. English is limited as described above; SSML, emotion and speed controls are not supported. Polyphonic characters, names and model numbers may require text adaptation. Other chips have not yet been adapted.
License
Examples, public headers and integration files use the MIT license. The precompiled engine, model, dictionary and demo firmware use the separate LingTTS Community License 1.0. The core and model are not open-source software. Personal projects, education, and internal company R&D, testing and prototypes are free; commercial deployment requires written authorization. Contact service@lingcorex.cn.
Report board, documentation or pronunciation issues via GitHub Issues, including the board, SDK version, memory configuration and relevant logs. Do not include secrets or personal information.