voice_audio_multi/docs/models/magpie_tts.md

MagpieTTS

MagpieTTS Multilingual 357M is wired as --family magpie_tts for offline text to speech. It uses baked speaker context prompts, a decoder transformer, an autoregressive local transformer, and NanoCodec waveform decoding.

Install

python3 tools/model_manager_v2.py install magpie_tts_orig

The default package installs a standalone GGUF directory:

models/MagpieTTS-Multilingual-357M-GGUF

Quick Start

audiocpp_cli --task tts --family magpie_tts \
  --model models/MagpieTTS-Multilingual-357M-GGUF \
  --backend cuda \
  --language en \
  --text "The production coordinator reviewed the overnight audio report and sent one clear update." \
  --voice-id Sofia \
  --out out.wav

German example:

audiocpp_cli --task tts --family magpie_tts \
  --model models/MagpieTTS-Multilingual-357M-GGUF \
  --backend cuda \
  --language de \
  --text "Das Team überprüfte am Morgen die Aufzeichnungen und bereitete einen kurzen Bericht vor." \
  --request-option voice_id=Jason \
  --out out_de.wav

Model

Field	Value
Family	magpie_tts
Task	tts
Modes	offline
Model directory	models/MagpieTTS-Multilingual-357M-GGUF
Languages	ar-AE, ar-MSA, ar-SA, de, en, es, fr, hi, it, ko, pt-BR, vi, zh
Voice input	Baked voice prompt through voice_id
Reference cloning	Not exposed by this integration

Language note: the upstream NVIDIA checkpoint advertises 12 languages, including Japanese (ja). The current audio.cpp integration exposes only the language frontends listed above. The standalone GGUF embeds the model config and sidecar resources, but Japanese requires NeMo's JapanesePhonemeTokenizer / JapaneseKatakanaAccentG2p path backed by OpenJTalk-style text processing. That tokenizer path has not been ported to the native C++ frontend yet, so ja is intentionally not listed as supported here.

voice_id accepts either a baked speaker name from the package speaker map or a numeric speaker index. --voice-id takes precedence over --request-option voice_id=...; both forms select the same baked voices.

Common Options (use directly)

Option	Values	Default	Meaning
--language	language code	en	Target synthesis language used by the Magpie text frontend.
--temperature	float >= 0	0.6	Audio-token sampling temperature.
--top-k	integer > 0	80	Audio-token top-k sampling limit.
--guidance-scale	float >= 0	2.5	Classifier-free guidance scale for decoder logits.
--max-tokens	integer > 0	500	Maximum decoder frames per text segment.
--text-chunk-size	integer > 0	100 or 300	Long-form text chunk budget: 100 for zh and ko, otherwise 300. Each chunk is generated within max_tokens.
--text-chunk-mode	default, tag_aware, japanese, endline	default	Framework text chunking mode.
--seed	integer >= 0	0	Audio-token sampling seed.

Request Options (use with --request-option)

Option	Values	Default	Meaning
language	language code	en	Target synthesis language used by the Magpie text frontend.
voice_id	Aria, Jason, John, Leo, Sofia, or numeric index	0	Packaged MagpieTTS voice name; default Aria. Numeric speaker index strings are also accepted within the checkpoint speaker count.
temperature	float >= 0	0.6	Audio-token sampling temperature.
top_k	integer > 0	80	Audio-token top-k sampling limit.
guidance_scale	float >= 0	2.5	Classifier-free guidance scale for decoder logits.
max_tokens	integer > 0	500	Maximum decoder frames per text segment.
text_chunk_size	integer > 0	100 or 300	Long-form text chunk budget: 100 for zh and ko, otherwise 300. Each chunk is generated within max_tokens.
text_chunk_mode	default, tag_aware, japanese, endline	default	Framework text chunking mode.
seed	integer >= 0	0	Audio-token sampling seed.

Session Options (use with --session-option)

Option	Values	Default	Meaning
magpie_tts.graph_arena_mb	integer MiB	1024	Reusable graph arena size for Magpie stages.
magpie_tts.weight_context_mb	integer MiB	2048	Weight metadata arena size.
magpie_tts.weight_type	native, f32, f16, bf16, q8_0	native	Matmul weight storage type when supported.
magpie_tts.conv_weight_type	native, f32, f16	native	Convolution weight storage type when supported.