voice_audio_multi/docs/models/personaplex.md

PersonaPlex

PersonaPlex is wired as --family personaplex for speech-to-speech conversation. It consumes user speech, applies a packaged or user-provided voice prompt, and generates an assistant speech response.

Install

python3 tools/model_manager_v2.py install personaplex_7b_v1_q4_k

The Q4_K package is the default. A Q8_0 package is also available:

python3 tools/model_manager_v2.py install personaplex_7b_v1_q8_0

Both packages install into:

models/PersonaPlex-GGUF

Quick Start

Packaged voice prompt:

audiocpp_cli --task s2s --family personaplex \
  --model models/PersonaPlex-GGUF \
  --backend cuda \
  --audio user.wav \
  --text "You are a concise assistant. Answer naturally and briefly." \
  --request-option voice_id=NATF2 \
  --out reply.wav

Raw reference voice:

audiocpp_cli --task s2s --family personaplex \
  --model models/PersonaPlex-GGUF \
  --backend cuda \
  --audio user.wav \
  --voice-ref assets/resources/a.wav \
  --text "You are a calm support specialist. Keep the caller reassured." \
  --out reply_ref.wav

Streaming mode:

audiocpp_cli --task s2s --mode streaming --family personaplex \
  --model models/PersonaPlex-GGUF \
  --backend cuda \
  --audio user.wav \
  --request-option voice_id=NATM1 \
  --text "You are a helpful assistant. Answer the user directly." \
  --out reply_stream.wav

Model

Field	Value
Family	personaplex
Task	s2s
Modes	offline, streaming
Model directory	models/PersonaPlex-GGUF
Input	User speech WAV through --audio
Prompt	System/persona prompt through --text or system_prompt
Voice input	Packaged voice_id or user WAV through --voice-ref
Language	English

If system_prompt is not provided, the runtime uses --text as the system prompt and wraps plain text with the tags expected by the model.

Packaged Voices

The package includes these voice prompt ids:

NATF0 NATF1 NATF2 NATF3
NATM0 NATM1 NATM2 NATM3
VARF0 VARF1 VARF2 VARF3 VARF4
VARM0 VARM1 VARM2 VARM3 VARM4

Use --request-option voice_id=<id> to select one. If omitted, NATF2 is used. If --voice-ref is supplied, the runtime uses that reference audio instead of a packaged prompt.

Common Options (use directly)

Option	Values	Default	Meaning
--audio	WAV path	required	User speech input.
--text	text	empty	Assistant persona/system prompt.
--voice-ref	WAV path	not set	User reference voice prompt. Overrides packaged voice_id when present.
--temperature	float >= 0	0.8	Audio-token temperature; must be > 0 when do_sample=true, ignored when false.
--top-k	integer >= 0	250	Audio-token top-k sampling limit.
--seed	integer >= 0	see description	uint32 seed. Session RNG starts at 42424242; explicit values reseed it, while omission on later conversations continues RNG state.

Request Options (use with --request-option)

Option	Values	Default	Meaning
system_prompt	text	empty	Assistant persona/system prompt.
voice_id	packaged id	NATF2	Packaged PersonaPlex voice prompt id.
temperature	float >= 0	0.8	Audio-token temperature; must be > 0 when do_sample=true, ignored when false.
text_temperature	float >= 0	follows temperature	Text-token temperature; must be > 0 when do_sample=true, ignored when false.
top_k	integer >= 0	250	Audio-token top-k sampling limit.
text_top_k	integer >= 0	follows top_k	Text-token top-k sampling limit.
do_sample	bool	true	Enable stochastic sampling. Set false for greedy decoding.
seed	integer >= 0	see description	uint32 seed. Session RNG starts at 42424242; explicit values reseed it, while omission on later conversations continues RNG state.

Session Options (use with --session-option)

Option	Values	Default	Meaning
personaplex.graph_arena_mb	integer MiB	1024	Reusable graph arena size for LM, depformer, and Mimi graphs.
personaplex.lm_weight_context_mb	integer MiB	64	Main LM weight metadata arena size.
personaplex.depformer_weight_context_mb	integer MiB	64	Depth transformer weight metadata arena size.
personaplex.mimi_weight_context_mb	integer MiB	64	Mimi codec weight metadata arena size.
personaplex.weight_type	native, f32, f16, bf16, q8_0	native	LM and Mimi matmul weight storage type when supported.