Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.fluidinference.com/llms.txt

Use this file to discover all available pages before exploring further.

​
Overview

PocketTTS (~155M params) is an autoregressive TTS backend that generates audio frame-by-frame. No espeak dependency — uses SentencePiece tokenization directly. Audio starts streaming ~80ms after prefill. Model: FluidInference/pocket-tts-coreml

​
Quick Start

import FluidAudioTTS

let manager = PocketTtsManager()
try await manager.initialize()

let audioData = try await manager.synthesize(text: "Hello, world!")

try await manager.synthesizeToFile(
    text: "Hello, world!",
    outputURL: URL(fileURLWithPath: "/tmp/output.wav")
)

​
Architecture

PocketTtsManager.synthesize(text:)
  → chunkText() — split into max 50 token chunks
  → loadMimiInitialState() — 23 streaming state tensors
  → FOR EACH CHUNK:
      → tokenizer.encode() — SentencePiece
      → embedTokens() — table lookup
      → prefillKVCache() — 125 voice + N text tokens
      → GENERATE LOOP:
          → runFlowLMStep() — transformer_out + eos_logit
          → flowDecode() — 8 Euler steps → 32-dim latent
          → denormalize() → quantize() → runMimiDecoder()
          → 1920 audio samples per frame
  → concatenate + postprocess
  → WAV output (24kHz mono)

​
Key State

​
KV Cache

  • 6 cache tensors [2, 1, 512, 16, 64] + 6 position counters
  • Reset per chunk

​
Mimi State

  • 23 tensors for convolution history, attention caches, overlap-add buffers
  • Continuous across chunks — keeps audio seamless

​
Text Chunking

Long text splits at 50 tokens or fewer:
  1. Sentence boundaries (.!?)
  2. Clause boundaries (,;:)
  3. Word boundaries (fallback)

​
Pipeline

text → SentencePiece tokenizer → subword tokens → PocketTTS model → audio
                                                    ↑
                                          pronunciation decisions
                                          happen inside model weights
                                          (no external control)
Unlike Kokoro which uses espeak to convert text to IPA phonemes before the model, PocketTTS feeds raw text tokens directly into the neural network. The model learned text→pronunciation mappings during training — there is no phoneme stage to intercept.

​
Pronunciation Control

FeatureSupportedWhy
SSML <phoneme>NoNo IPA layer — model has no phoneme vocabulary
Custom lexicon (word → IPA)NoNo phoneme stage to apply mappings
Markdown [word](/ipa/)NoSame — no phoneme input
SSML <sub> (text substitution)PlannedText-level, can run before tokenizer
Text preprocessing (numbers, dates)PlannedText-level, can run before tokenizer
What can be added — anything that operates on text before the SentencePiece tokenizer: number/date/currency expansion, text substitution, abbreviation expansion. What cannot be added without retraining — anything that requires phoneme-level control. The model decides pronunciation from text tokens alone. See Kokoro if you need pronunciation control.

​
CoreML Details

  • All 4 models loaded with .cpuAndGPU (ANE float16 causes artifacts in Mimi state)
  • Compiled from .mlpackage → .mlmodelc on first load, cached on disk
  • Thread-safe via actor pattern

​
Benchmarks

Benchmarks in progress. Methodology follows Kyutai’s evaluation and their tts_longeval toolkit.

​
Upstream (Kyutai, CPU)

LibriSpeech test-clean, WER via Whisper large-v3:
MetricPocketTTS (100M)F5-TTSDSM (313M)
WER1.84%2.21%1.84%
Audio Quality (ELO)2016——
Speaker Similarity (ELO)1898——
Runs on CPUYes (6x real-time)NoNo
ELO from human pairwise evaluation (50 raters, 50 samples). Tested on Apple M3 and Intel Core Ultra 7.

​
FluidAudio CoreML (planned)

We will benchmark the CoreML port against the upstream PyTorch CPU baseline using the same methodology:
MetricHowDataset
WERTranscribe TTS output with Whisper large-v3, compare to input textLibriSpeech test-clean
Speaker SimilarityWavLM cosine similarity between prompt audio and generated audioLibriSpeech test-clean
RTFxWall-clock generation time / audio durationVariable length (1s to 300s)
Time to First AudioTime from synthesize() call to first audio frameSingle sentence
Peak RAMInstruments / os_proc_memory during generationVariable length
Additional datasets from tts_longeval:
  • NTREX — monologue sentences from news translation corpus
  • Synthetic Dialogs — daily life, technical, and number-heavy scripts
  • SEED English — adapted from ByteDance’s SEED TTS Eval
Key comparisons: CoreML ANE vs PyTorch CPU (upstream), CoreML vs Kokoro CoreML (FluidAudio internal).

​
License

CC-BY-4.0, inherited from kyutai/pocket-tts.