KingNish/chotu-v0.1

Chotu v0.1

A sub-2 MB streaming English voice that starts speaking in 30 ms on a single CPU thread.

Chotu v0.1 is an INT8 ONNX text-to-speech model built for browser runtimes, offline edge devices, and real-time voice agents. It streams 24 kHz audio chunk-by-chunk with persistent state and zero cloud calls.

Highlights

• Under 2 MB: Single INT8 ONNX graph (paradee_v4_1_int4.onnx, packaged as paradee_v4_1_int4.zip).
• 30 ms median TTFA: Emits the first 150 ms audio chunk immediately instead of waiting for full-utterance synthesis.
• 80x real-time on one CPU thread: Runs locally in-browser via onnxruntime-web (WASM SIMD) or natively via ONNX Runtime.
• Continuous-depth decoding: Tunable num_dec_loops lets callers trade compute for fidelity per request.

Audio Samples

Generated with Chotu v0.1 (24000 Hz, num_dec_loops = 4, speed = 1.0):

Sample	Text	Audio
Intro	"Hi, I am Chotu, an under two megabyte voice model streaming right in your browser."
Latency	"With thirty millisecond first audio latency and eighty x real time speed, Chotu speaks on a single CPU thread."
Edge	"Built for offline edge devices and real time voice agents, Chotu runs locally with zero cloud calls."

Benchmarks & Specs

Property	Value
Footprint	<2 MB (INT8 ONNX)
Sample Rate	24000 Hz (16-bit mono PCM)
Median TTFA	30 ms
Throughput	80x real-time (single CPU thread)
Vocab Size	178 tokens (config.json)
License	Apache 2.0

Streaming Timeline & Decoder Loop

Architecture

Chotu v0.1 runs as a single stateful ONNX graph across sliding token windows (64 active tokens with past context and lookahead), carrying KV, F0, ConvNeXt, and harmonic phase state between steps.

Streaming ONNX I/O

Inputs (12)

Tensor	Type	Shape	Description
input_ids	int64	[1, T_win]	Windowed token IDs from config.json
speed	float32	[1]	Speaking rate (1.0 = normal)
num_dec_loops	int64	[1]	Decoder refinement loops
tok_start	int64	[1]	Active window start token index
tok_end	int64	[1]	Active window end token index
frame_start	int64	[1]	Chunk start frame index
frame_end	int64	[1]	Chunk end frame index
kv1_cache_in	float32	[2, 1, 2, W_past, 32]	Encoder KV state 1
kv2_cache_in	float32	[2, 2, 1, 2, W_past, 32]	Encoder KV state 2
conv_cache_in	float32	[36, 1, 4, 256]	Causal ConvNeXt decoder state
f0_prev2_in	float32	[1, 4]	Recent F0 history
c0_in	float32	[1]	Harmonic phase state

Outputs (7)

Tensor	Type	Shape	Description
waveform	float32	[samples]	24 kHz PCM audio chunk
f0_hz	float32	[frames]	Predicted F0 contour
dur_frames	int64	[T_act]	Predicted frame durations per token
kv1_cache_out	float32	[2, 1, 2, W_out, 32]	Updated KV state 1
kv2_cache_out	float32	[2, 2, 1, 2, W_out, 32]	Updated KV state 2
conv_cache_out	float32	[36, 1, 5, 256]	Updated ConvNeXt state
c_subframes	float32	[subframes]	Phase state for next chunk

Usage

Coming soon.

Credits

Chotu v0.1 builds on Paradee-8M-v1.0 and Kokoro-82M.