aufklarer/Pocket-TTS-100M-CoreML-INT8

Pocket TTS 100M CoreML INT8

Native CPU speech synthesis for Apple platforms, with an English Alba preset. The language-model linear weights use the existing upstream INT8 conversion; conditioning, the eight-step fused flow decoder and Mimi retain FP16 weights. The bundle emits 1,920-sample frames at 24 kHz (80 ms) during generation.

Property	Value
Model	Kyutai Pocket TTS, 100M
Format	Precompiled CoreML .mlmodelc
Language / voice	English / Alba
Output	Mono Float32 PCM, 24 kHz
Cache	Six layers, 512 positions
Flow steps	Eight, fixed in the exported graph
Sampling	Temperature 0.7, request-local deterministic seed
Bundle	Approximately 299 MiB; exact sizes in manifest.json

Files

Path	Size	Purpose
cond_prefill.mlmodelc/	127.3 MiB	Text conditioning, FP16 weights
flowlm_stepv2.mlmodelc/	73.5 MiB	Autoregressive backbone, INT8 linear weights
flow_decoder_fused.mlmodelc/	18.5 MiB	Fused flow-matching decoder, FP16 weights
mimi_decoder.mlmodelc/	39.9 MiB	Streaming audio decoder, FP16 weights
constants_bin/	39.7 MiB	Tokenizer, text/BOS embeddings, padded Alba voice state
config.json	0.7 KiB	Versioned Swift runtime contract
manifest.json	5.6 KiB	SHA-256 and byte size for every packaged file
NOTICE, LICENSE	18.5 KiB	Attribution and CC BY 4.0 license

Performance

Measurements use the release speech-swift runtime with every graph on CPU. Synthesis timings exclude loading; RTF is wall time divided by output duration. Warmup and measured iterations are recorded separately. RSS and physical footprint are process memory metrics, not model file sizes.

Metric	Apple M5 Pro CPU-only
First PCM, median of five warm requests	78.9 ms
Full synthesis, median	943.4 ms / 2.8 s audio
RTF (lower is faster)	0.337
Peak process RSS, including load	866.8 MiB
Largest post-run physical footprint	343.7 MiB
Independent ASR round trip	0 / 150 word errors, 12 phrases

See the CPU benchmark report for measurements, round trips, and reproducible commands. These measurements do not establish listener preference or performance on an iPhone.

Usage

Use the Pocket TTS runtime on the feat/pocket-tts branch until it is merged into the default branch.

import PocketTTS

let model = try await PocketTTSModel.fromPretrained()
try await model.generateStreaming(text: "Hello world.") { chunk in
    await consume(chunk.samples)
}

speech speak "Hello world." --engine pocket --stream -o hello.wav
speech speak "Hello world." --engine pocket --stream \
  --pocket-bundle-dir /path/to/bundle --pocket-seed 42 -o hello.wav

The Python benchmark driver invokes the native release binary, then runs ASR in a separate process after synthesis exits. Python is not needed for Swift inference.

import json
import subprocess

run = subprocess.run([
    "pocket-tts-bench", "--bundle-dir", "/path/to/bundle",
    "--text", "Hello world.", "--runs", "3", "--warmup", "1",
    "--output-dir", "/tmp/pocket-bench",
], check=True)
with open("/tmp/pocket-bench/metrics.jsonl") as source:
    measurements = [json.loads(line) for line in source]

Limits

English and the fixed Alba preset only. No arbitrary voice cloning, speed control or multilingual pack loading. Text is limited to 4,096 characters and split around a 50-token target. A generation/cache limit throws instead of returning success with a clipped final word. Each successful stream ends with one empty final marker. Cancellation and failures terminate the stream with an error; previously delivered audio is partial.

Source and attribution

• Kyutai Pocket TTS
• CoreML conversion by FluidInference, revision 91748676fe3c8b2eb3007b3125253bcd898202c3, v2.1/english/.
• Alba voice by Alba MacKenna.

Weights and the Alba voice are CC BY 4.0. Retain Kyutai, FluidInference and Alba MacKenna attribution when redistributing. The learned graph weights are preserved, not requantized. Mimi omits redundant passthrough outputs to avoid CoreML BNNS input/output name collisions; the host retains that unchanged state. All graph operations and recurrent inputs are preserved. The voice snapshot is expanded into explicit padded F32 cache tensors for the native runtime.

Links

• speech-swift
• Installation and CLI docs
• Pocket TTS guide
• Speech Studio
• soniqo.audio
• Blog