voice_audio_multi/docs/models/pocket_tts_small.md

PocketTTS Small

pocket_tts_small is the dedicated English PocketTTS Small family. It supports text-to-speech (tts) and voice cloning (clon) in offline and streaming modes, using a dedicated implementation.

Packages

F32 packages are hosted under PocketTTS-Small-GGUF in the audio-cpp/audio.cpp-gguf repository. The default is 18m ReFL.

Package	GGUF file
pocket_tts_small_english_18m_refl_f32 (default)	english-18m-refl-f32.gguf
pocket_tts_small_english_18m_f32	english-18m-f32.gguf
pocket_tts_small_english_7m_refl_f32	english-7m-refl-f32.gguf
pocket_tts_small_english_7m_f32	english-7m-f32.gguf
pocket_tts_small_english_2m_refl_f32	english-2m-refl-f32.gguf
pocket_tts_small_english_2m_f32	english-2m-f32.gguf

Each package includes embeddings/alba.safetensors as its predefined alba voice. GGUF sidecars provide config.yaml and tokenizer.model; the configuration is required to describe the selected small model. The embedding remains beside the GGUF under embeddings/.

The source layout follows PocketTTS: GGUF weights at the model root with the configuration and tokenizer resolved through sidecars, or a Safetensors source containing model.safetensors, config.yaml, and tokenizer.model at the model root, with predefined voices under embeddings/. The six managed packages use GGUF.

Controls

The request controls match PocketTTS: frames_after_eos, max_frames, text_chunk_tokens, temperature, seed, noise_clamp, eos_threshold, noise_file, voice_embedding_file, truncate_clone_audio, text_chunk_size, and export_voice_state_file. Defaults and ranges are recorded in model_specs/pocket_tts_small.json.

Session controls use the pocket_tts_small. namespace, including pocket_tts_small.weight_type, pocket_tts_small.voice_state_cache_slots, pocket_tts_small.offline_graph_capacity_mode, and the corresponding PocketTTS flow/Mimi storage and graph capacity controls. Use --session-option name=value for session controls and --request-option name=value for request controls. There are no public family or option aliases and no language load selector; these packages are English.

Select --family pocket_tts_small, --task tts for a predefined voice or --task clon with speaker reference audio. Both tasks accept --mode offline and --mode streaming. Enable --log and save the log when running either mode.

Quick start

audiocpp_cli --family pocket_tts_small --task tts \
  --model models/PocketTTS-Small-GGUF/english-18m-refl/english-18m-refl-f32.gguf \
  --backend cpu --voice-id alba \
  --text "Hello! This is the small PocketTTS model speaking locally." \
  --out output.wav --log

Use --mode streaming for streaming synthesis. For voice cloning, select --task clon and pass --voice-ref reference.wav instead of --voice-id alba.

The checkpoints come from Kyutai PocketTTS and use the CC-BY-4.0 license. The implementation follows upstream revision 4cc57533.

Offline synthesis uses bounded causal blocks of four latent frames by default on every backend. --session-option pocket_tts_small.use_full_mimi=true selects full-sequence decoding explicitly.

Experimental native CPU flow

The default implementation uses GGML on every backend. A build with -DENGINE_ENABLE_NEON=ON also provides an experimental F32 CPU flow implementation. The audio decoder continues to use GGML. Select it explicitly with:

--session-option 'pocket_tts_small.exec_plan={"flow":{"impl":"native","params":{"threads":1}}}'

The plan's threads parameter controls native operators; the CLI --threads option controls the GGML decoder. Performance depends on model size and CPU: benchmark both implementations for your workload. Native flow requires F32 weights and a CPU backend. GPU backends use GGML.

Experimental offline batching

C++ callers can use runtime::IBatchedOfflineVoiceTaskSession::run_batch with one TaskRequest per utterance. Results retain input order. Each request can select its own VoiceCondition and override generation options such as seed, temperature, and max_frames; omitted values use the prepared session defaults. Each row has independent random state, cache positions, and EOS stopping. Long requests retain sentence chunking and concatenate their chunks into one result.

Single-request, streaming, and batched generation share the same GGML acoustic decode graph implementation; ordinary generation uses batch size one. Text prefill and Mimi waveform decoding remain per-request. Graphs belong to the session and are reused while their batch/cache capacities fit. The existing graph-capacity policy also applies to batch cache sizing. Finished rows retain their slots until the other rows finish, so mixed lengths can limit throughput. Memory grows with batch size and cache capacity; a larger batch is not always faster.

Batching requires GGML flow and offline mode. Streaming and native CPU flow continue to use the single-request interface. Batched floating-point execution can change sampled audio; identical seeds do not promise identical WAVs across batch sizes or compositions. Native CPU flow remains a separate implementation.

The model integration test pocket_tts_small_batch exercises the batch interface. The CLI's --batch-text-file schedules requests one at a time through that same batch-size-one runtime; C++ callers use the batch interface to group requests.

Performance controls

pocket_tts_small.attention (auto, eager, or flash) and pocket_tts_small.kv_cache_type (f32 or f16) independently control acoustic attention and KV storage. Defaults are automatic attention selection and F32 KV storage. These session options apply equally to ordinary, streaming, and batched GGML generation. Native CPU flow requires eager attention and F32 state.

For example, add --session-option pocket_tts_small.attention=flash and --session-option pocket_tts_small.kv_cache_type=f16 to the CLI command. These controls do not change the GGUF weights. Reduced precision and attention changes can change sampled speech; compare quality on your workload.

Waveform decoding defaults to bounded causal blocks of four latent frames. pocket_tts_small.mimi_chunk_frames controls that block size; larger values can improve throughput but use more memory. Graphs are reused between blocks and utterances, while convolution history and transformer state reset for each new utterance. The final block is padded and only its valid audio is returned. pocket_tts_small.use_full_mimi=true selects a full-utterance graph instead, with a higher peak-memory cost. Both paths implement the same causal decoder.