README.md · sanoTTS

1

---

2

license: gpl-3.0

3

language: [en, ne, hi, vi, id, zh]

4

pipeline_tag: text-to-speech

5

tags: [text-to-speech, tiny, microcontroller, esp32, wasm, piper, distillation]

6

library_name: sanotts

7

---

8

9

# sanoTTS — a tiny neural voice that runs anywhere

10

11

***sano*** (सानो) — Nepali for **"small."** A family of tiny neural

12

text-to-speech voices — **294k to 2.27M parameters** — that run with

13

**no cloud and no NPU**: real-time on a ~$3 ESP32-S3 (out a GPIO into an

14

LM386 and a speaker), or live in the browser via WASM.

15

16

![text in → ESP32 → speech out](https://raw.githubusercontent.com/Ampixa/sanoTTS/master/docs/assets/saanotts-mcu-hero.png)

17

18

| | |

19

| --- | --- |

20

| Parameter range | **294,279 – 2,272,145** |

21

| Runs real-time on a $3 microcontroller | ESP32-S3, out a GPIO into an LM386 |

22

| Runs in the browser | WebAssembly, no server |

23

| Per-voice footprint | **337 KB to 8.7 MB**, zero dependencies (espeak-ng phonemizer included) |

24

| Coverage | **11 voices** across **6 languages** — English, Nepali (नेपाली), Hindi (हिन्दी), Vietnamese (Tiếng Việt), Indonesian (Bahasa), Chinese (中文) |

25

| License | open source, **GPL-3.0** |

26

27

**Live demo:** [tts.ampixa.com/sanoTTS](https://tts.ampixa.com/sanoTTS) —

28

every voice synthesizes your text live in the browser, no server, no upload.

29

30

## Download

31

32

Both packages stream their weights from **this repo** by default.

33

34

```python

35

import sanotts

36

sanotts.synthesize("Hello world", voice="heart") # downloads heart/ from here

37

```

38

39

```js

40

import { SanoTTS } from 'sanotts-web'; // fetches web/voices/ from here

41

const tts = await SanoTTS.load();

42

const result = await tts.synthesize('Hello world', { voice: 'heart' });

43

```

44

45

Python needs `sanotts` >= 0.3.0, the browser `sanotts-web` >= 0.3.0. Both fall

46

back to the GitHub releases or the Pages host if Hugging Face cannot be

47

reached, so an outage here cannot break an install. Python packages land in

48

`~/.cache/sanotts/`; set `SANOTTS_VOICE_SOURCE=hf` or `=github` to pin one

49

host. In the browser, passing `voiceBase` yourself turns the fallback off, so

50

a self-hosted deployment never quietly reaches back to our servers.

51

52

## Samples

53

54

Two clips per voice are in this repo's `samples/` folder; one is embedded

55

below. "Package here" means this repo carries that voice's weights. The three

56

languages without one currently ship only through the

57

[browser demo](https://tts.ampixa.com/sanoTTS) and the GitHub repo's

58

`web/voices/` — their packages haven't been exported yet.

59

60

61

| --- | --- | ---: | :---: | :---: | --- |

62

| heart | English 🇺🇸 | 2.27 M | 3.48 | [`heart/`](https://huggingface.co/ampixa/sanoTTS/tree/main/heart) | <audio controls src="https://huggingface.co/ampixa/sanoTTS/resolve/main/samples/english-heart-1.mp3"></audio> |

63

| hfc | English 🇺🇸 | 1.83 M | 3.94 | [`hfc-en-1p8m/`](https://huggingface.co/ampixa/sanoTTS/tree/main/hfc-en-1p8m) | <audio controls src="https://huggingface.co/ampixa/sanoTTS/resolve/main/samples/english-hfc-1.mp3"></audio> |

64

| amy | English 🇺🇸 | 1.46 M | **4.13** | [`amy-en-1p46m/`](https://huggingface.co/ampixa/sanoTTS/tree/main/amy-en-1p46m) | <audio controls src="https://huggingface.co/ampixa/sanoTTS/resolve/main/samples/english-amy-1.mp3"></audio> |

65

| kristin | English 🇺🇸 | 1.40 M | 4.09 | [`kristin-en-1p4m/`](https://huggingface.co/ampixa/sanoTTS/tree/main/kristin-en-1p4m) | <audio controls src="https://huggingface.co/ampixa/sanoTTS/resolve/main/samples/english-kristin-1.mp3"></audio> |

66

| amy-small | English 🇺🇸 | 1.08 M | 3.70 | [`amy-en-1p1m/`](https://huggingface.co/ampixa/sanoTTS/tree/main/amy-en-1p1m) | <audio controls src="https://huggingface.co/ampixa/sanoTTS/resolve/main/samples/english-amy-small-1.mp3"></audio> |

67

| robot (on-device, int8) | English 🇺🇸 | 567 k | — | not packaged here (int8 MCU format) | <audio controls src="https://huggingface.co/ampixa/sanoTTS/resolve/main/samples/mcu-567k.mp3"></audio> |

68

| heart-nano | English 🇺🇸 | 294 k | 2.29 | [`heartnano/`](https://huggingface.co/ampixa/sanoTTS/tree/main/heartnano) | <audio controls src="https://huggingface.co/ampixa/sanoTTS/resolve/main/samples/english-heart-nano-1.mp3"></audio> |

69

| Indonesian | Bahasa | 1.46 M | — | [`id-newstts-1p46m/`](https://huggingface.co/ampixa/sanoTTS/tree/main/id-newstts-1p46m) | <audio controls src="https://huggingface.co/ampixa/sanoTTS/resolve/main/samples/indonesian-1.mp3"></audio> |

70

| Vietnamese | Tiếng Việt | 1.46 M | — | [`vi-vais1000-1p46m/`](https://huggingface.co/ampixa/sanoTTS/tree/main/vi-vais1000-1p46m) | <audio controls src="https://huggingface.co/ampixa/sanoTTS/resolve/main/samples/vietnamese-1.mp3"></audio> |

71

| Nepali | नेपाली | 1.47 M | — | not exported yet — see `web/voices/nepali/` in the GitHub repo | <audio controls src="https://huggingface.co/ampixa/sanoTTS/resolve/main/samples/nepali-1.mp3"></audio> |

72

| Hindi | हिन्दी | 1.50 M | — | not exported yet — see `web/voices/hindi/` in the GitHub repo | <audio controls src="https://huggingface.co/ampixa/sanoTTS/resolve/main/samples/hindi-1.mp3"></audio> |

73

| Chinese | 中文 | 1.50 M | — | not exported yet — see `web/voices/chinese/` in the GitHub repo | <audio controls src="https://huggingface.co/ampixa/sanoTTS/resolve/main/samples/chinese-1.mp3"></audio> |

74

75

SCOREQ is a no-reference naturalness predictor, higher is better; it is only

76

reported for the English voices, which share one 24-sentence eval set. The

77

other languages haven't been scored against a comparable reference yet.

78

79

**Size does not order this table, and we are not going to pretend it does.**

80

`amy` at 1.46M scores higher than `heart` at 2.27M because they come from

81

different teachers and different architectures, not because one is bigger.

82

`amy` and the voices above it are the **piperlite** lineage, distilled from a

83

Piper/VITS teacher at 22.05 kHz. `heart` and `heart-nano` are the newer

84

**nano** lineage — 24 kHz, mel-100 → ConvNeXt1D → iSTFT, distilled from a

85

Kokoro teacher through a frozen Vocos — which is the one that quantizes to

86

int8 and runs on a microcontroller. The two are not interchangeable and the

87

`sanotts` package picks the right runtime for you.

88

89

The `heart` and `heart-nano` scores were re-measured on 2026-09-04; the eval

90

set, checkpoint hashes, exact commands and all 24 per-clip scores are in

91

[`evidence/heart-diverse24-remeasure-20260904.json`](https://huggingface.co/ampixa/sanoTTS/blob/main/evidence/heart-diverse24-remeasure-20260904.json).

92

They were scored on the float32 reference render. The shipped `heart-nano`

93

int8 export tracks that render at 0.981 waveform correlation, and `heart`

94

ships as float32 because its int8 export reached only 0.951 against a 0.98

95

gate.

96

97

The "robot" row is the 567,008-parameter model that runs on the ESP32-S3.

98

99

## Install & use

100

101

| Platform | Install | Then |

102

| --- | --- | --- |

103

| Python | `pip install sanotts` | `sanotts say "Hello" --voice heart -o hello.wav` |

104

| Web (npm) | `npm install sanotts-web` | `const tts = await SanoTTS.load(); await tts.synthesize('Hello', {voice:'heart'})` |

105

| Web (no build) | copy `dist/` + `voices/` | see [Deploy on your own site](https://github.com/Ampixa/sanoTTS#deploy-on-your-own-site) in the GitHub README |

106

| Arduino / PlatformIO | zip-install or `lib_deps = https://github.com/Ampixa/sanoTTS.git` | [`arduino/README.md`](https://github.com/Ampixa/sanoTTS/blob/master/arduino/README.md) |

107

| Hugging Face | this repo | the voice packages above, downloaded for you by `pip install sanotts` |

108

| Browser | nothing | [tts.ampixa.com/sanoTTS](https://tts.ampixa.com/sanoTTS) |

109

110

Pip voices: `heart`, `hfc`, `amy-1p8m`, `amy`, `kristin`, `vi`, `id`,

111

`amy-1p1m`, `heart-nano`. Pure numpy inference, no torch, no onnxruntime.

112

113

## How it stacks up

114

115

Open small-scale TTS on an honest gate — a diverse 24-sentence set scored

116

with the **same** no-reference suite (SCOREQ / UTMOS are naturalness

117

predictors, DNSMOS-SIG is signal quality; higher is better). Parameter

118

counts are inference-time and exclude the shared external G2P.

119

120

121

| --- | ---: | :---: | :---: | :---: |

122

| **sanoTTS (amy)** | **1.46 M** | **4.13** | **4.10** | 3.61 |

123

| TinyTTS | 1.62 M | 3.94 | 3.65 | **3.62** |

124

| Inflect Nano | 4.63 M | 3.81 | 3.65 | 3.58 |

125

| Kitten TTS nano | 15 M | 3.02 | 3.58 | 3.43 |

126

| _Piper (our teacher)_ | _~15 M_ | _4.71_ | _4.47_ | _3.65_ |

127

| _Kokoro_ | _82 M_ | _4.89_ | _4.52_ | _3.69_ |

128

129

sanoTTS is the **smallest** model here and the **best on naturalness

130

(SCOREQ and UTMOS) among everything up to 15M params** — beating TinyTTS

131

while being smaller. On DNSMOS-SIG, TinyTTS edges us by 0.01 — no single

132

metric tells the whole story. It's the only one that runs a full neural

133

stack on a $3 MCU. The frontier only pulls ahead at Piper, the ~15M teacher

134

we distilled from, and Kokoro (82M, 56× larger) — a gap we don't claim to

135

close. Reproduce it with `tools/eval_mos_all.py` + `tools/eval_scorecard.py`

136

in the GitHub repo.

137

138

Shipped-file sizes: sanoTTS amy 2.8 MB fp16 and TinyTTS 3.5 MB fp16, both

139

verified from the released files; Kokoro's ~330 MB fp32 is its widely cited

140

public figure.

141

142

## How it works

143

144

![text → duration → acoustic → decoder → audio](https://raw.githubusercontent.com/Ampixa/sanoTTS/master/docs/assets/saanotts-signal-path.png)

145

146

espeak-ng provides phoneme IDs; a duration model predicts timing; an

147

acoustic model predicts generator latents; a decoder renders audio. The

148

piperlite voices (amy, kristin, hfc, and the other languages) use a compact

149

time-domain decoder running in fp32 at 22.05 kHz. The nano voices (heart,

150

heart-nano) and the 567,008-parameter on-device model use an iSTFT decoder

151

at 24 kHz, quantized to int8 where it has to fit and run in real time on the

152

ESP32-S3. Models are distilled from a Piper/VITS or Kokoro teacher — see

153

[`docs/distillation-recipe.md`](https://github.com/Ampixa/sanoTTS/blob/master/docs/distillation-recipe.md)

154

in the GitHub repo for the full recipe.

155

156

## Deploy

157

158

- **ESP32-S3 talking device** — a standalone WiFi dashboard: type text, the

159

board phonemizes (on-chip espeak-ng) and speaks. See

160

[`mcu/ports/esp32s3/`](https://github.com/Ampixa/sanoTTS/tree/master/mcu/ports/esp32s3).

161

Board-by-board measurements are in

162

[`BOARDS.md`](https://github.com/Ampixa/sanoTTS/blob/master/BOARDS.md); the

163

silicon figures there are for the `en_us_e12nano` lineage, which is a

164

sibling of `heart-nano`, not the same weights.

165

- **Browser** — the full stack in WASM, no server. **[▶ Hear and synthesize

166

all 11 voices live](https://tts.ampixa.com/sanoTTS)**; source in

167

[`web/`](https://github.com/Ampixa/sanoTTS/tree/master/web).

168

- **Other MCUs** — which chips can run it and how well:

169

[`docs/mcu-classes-and-porting.md`](https://github.com/Ampixa/sanoTTS/blob/master/docs/mcu-classes-and-porting.md).

170

171

## Links

172

173

- Source, recipes, eval tooling: [github.com/Ampixa/sanoTTS](https://github.com/Ampixa/sanoTTS)

174

- Live browser demo: [tts.ampixa.com/sanoTTS](https://tts.ampixa.com/sanoTTS)

175

- npm package: [sanotts-web](https://www.npmjs.com/package/sanotts-web)

176

- PyPI package: [sanotts](https://pypi.org/project/sanotts/)

177

178

## License

179

180

GPLv3 — see [`LICENSE`](https://github.com/Ampixa/sanoTTS/blob/master/LICENSE).

181

The pipeline builds on GPLv3 components (notably

182

[espeak-ng](https://github.com/espeak-ng/espeak-ng) for G2P, and

183

[piper](https://github.com/OHF-Voice/piper1-gpl)), so the project as a

184

whole is GPLv3.

185

186

187

188

## Files here

189

190

Two package layouts, because there are two graphs.

191

192

**piperlite** (`amy-en-1p46m/`, `kristin-en-1p4m/`, `hfc-en-1p8m/`,

193

`amy-en-1p1m/`, `id-newstts-1p46m/`, `vi-vais1000-1p46m/`) — a flat fp16

194

blob addressed by manifest offsets:

195

196

```

197

manifest.json weights.fp16.bin

198

piper-phoneme-config.json (+ sibilant-injection calibration where applicable)

199

```

200

201

**nano** (`heart/`, `heartnano/`) — the mel-100 stack, two blobs plus the

202

generated offset header:

203

204

```

205

meta.json lineage, per-file sha256, sample rate, vocab

206

front_*.bin duration + acoustic

207

model_*.bin decoder

208

nano_q8_meta.h tensor offsets

209

```

210

211

`heartnano/` ships `*_q8.bin` (int8, 345,232 bytes total); `heart/` ships

212

`*_f32.bin` (float32, 9,137,920 bytes) because its int8 export missed the

213

fidelity gate. Both layouts are consumed by the `sanotts` Python package and

214

the portable C runtime.

215

216

**`web/voices/`** is a third thing and exists for the browser. The piperlite

217

voices ship there as `front_f32.bin` + `dec_f32.bin`, a different artifact

218

from the `weights.fp16.bin` the Python package reads, which is why the same

219

voice appears twice in this repo under two names. It mirrors `web/` in the

220

GitHub repo byte for byte and is what `sanotts-web` fetches. The nano voices

221

are not duplicated: `web/voices/heart/` and `heart/` hold the same blobs.

222

223

`samples/` holds the audio clips embedded above (mp3, one or two per voice,

224

plus the on-device `mcu-567k.mp3`), and `evidence/` the eval report behind the

225

heart scores.

226