Skip to content

Correct OuteTTS model_specs_v1 metadata #124

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
0xShug0 merged 2 commits into from
Jul 27, 2026
+16 −10
Merged

Correct OuteTTS model_specs_v1 metadata #124

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
ed7f3f4
Update OuteTTS model spec v1 metadata
mirek190 Jul 27, 2026
a536314
Use normalized OuteTTS session options
mirek190 Jul 27, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 5 additions & 2 deletions docs/community_models/outetts.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -45,7 +45,7 @@ audiocpp_cli --task clon --family outetts \
--max-tokens 1024 --out cloned.wav
```

`--task tts` with the same `--voice-ref` and `--reference-text` options also enables speaker conditioning, which is useful for clients that expose one TTS route. Safetensors packages and older OuteTTS GGUFs do not contain the aligner. For those models, pass `outetts.aligner_model_path`; cloning fails clearly instead of using unreliable estimated word boundaries.
`--task tts` with the same `--voice-ref` and `--reference-text` options also enables speaker conditioning, which is useful for clients that expose one TTS route. Safetensors packages and older OuteTTS GGUFs do not contain the aligner. For those models, pass `outetts.aligner_path`; cloning fails clearly instead of using unreliable estimated word boundaries.

The installer places `DAC.speech.v1.0` and `Qwen3-ForcedAligner-0.6B` beside the OuteTTS directory. It converts the official DAC checkpoint to a safe tensor source. To do that conversion manually:

Expand Down Expand Up @@ -97,7 +97,10 @@ audiocpp_cli --task tts --family outetts \
| `--reference-text` | text | none | Exact transcript of `--voice-ref`; required for voice cloning. |
| `--request-option reference_language=<code>` | language code | `en` | Language used by the optional reference aligner. |
| `--session-option outetts.weight_type=native|f32|f16|bf16|q8_0` | enum | `native` | Language-model weight storage type. For CUDA voice cloning, quantized weights remain compact in the GGUF but are expanded to F32 in VRAM to avoid generation divergence over long reference-codec prompts. Normal TTS and CPU cloning keep the selected type. |
| `--session-option outetts.aligner_model_path=<path>` | model path | embedded aligner | Optional external Qwen3 Forced Aligner override, required only for safetensors packages and older GGUFs without the embedded aligner. |
| `--session-option outetts.aligner_path=<path>` | model path | embedded aligner | Optional external Qwen3 Forced Aligner override, required only for safetensors packages and older GGUFs without the embedded aligner. |

The legacy session keys `outetts.aligner_model_path` and
`outetts.dac_graph_context_mb` remain accepted for backward compatibility.
| `--session-option outetts.reference_cache_slots=<n>` | integer | `1` | LRU slots for prepared reference profiles (alignment, DAC codes, and word features). Set `0` to disable reuse. |
| `--session-option outetts.mem_saver=true|false` | bool | `false` | Release the reusable Llama cached-step graph after each generated chunk and release the aligner runtime after preparing a reference. Model and DAC weights stay resident; later requests rebuild released state. |

Expand Down
2 changes: 1 addition & 1 deletion docs/reports/outetts_validation.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -216,7 +216,7 @@ build\windows-cuda-release\bin\outetts_warm_bench.exe `
--backend cuda --threads 8 `
--request-file tests\outetts\warm_bench_requests.json `
--session-option outetts.weight_type=f32 `
--session-option outetts.aligner_model_path=..\models\Qwen3-ForcedAligner-0.6B `
--session-option outetts.aligner_path=..\models\Qwen3-ForcedAligner-0.6B `
--audio-out-dir build\reference\cpp_f32_final_explicit `
--log-file build\reference\cpp_f32_final_explicit.log
```
Expand Down
3 changes: 2 additions & 1 deletion model_specs_v1/outetts.json
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -272,11 +272,12 @@
"default": true,
"format": "gguf",
"precision": "q8_0",
"target_directory": "Llama-OuteTTS-1.0-1B-GGUF",
"target_directory": "Llama-OuteTTS-1.0-1B_Q8",
"files": [
"Text to audio (TTS)/Llama-OuteTTS-1.0-1B_Q8.gguf"
],
"strip_prefix": "Text to audio (TTS)",
"description": "Standalone Q8_0 GGUF with the IBM DAC codec and Qwen3 Forced Aligner weights, tokenizers, configs, and sidecars embedded for one-file TTS and voice cloning.",
"download": {
"kind": "huggingface_snapshot",
"repo": "mirek190/audio.cpp"
Expand Down
6 changes: 3 additions & 3 deletions src/community_models/outetts/loader.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -75,9 +75,9 @@ runtime::ModelCliInterface cli(const OuteTTSAssets &) {
"Language-model constant tensor context size in MiB."},
{"outetts.dac_weight_context_mb", "n",
"DAC decoder weight context size in MiB."},
{"outetts.dac_graph_context_mb", "n",
"DAC decoder graph context size in MiB."},
{"outetts.aligner_model_path", "path",
{"outetts.dac_graph_arena_mb", "n",
"DAC decoder graph arena size in MiB."},
{"outetts.aligner_path", "path",
"Optional Qwen3 Forced Aligner override. Cloning automatically uses "
"the aligner embedded in a standalone OuteTTS GGUF when present."},
{"outetts.reference_cache_slots", "n",
Expand Down
8 changes: 5 additions & 3 deletions src/community_models/outetts/session.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -438,7 +438,8 @@ resolve_aligner_assets(const runtime::SessionOptions &options,
const OuteTTSAssets &model_assets) {
const auto model_path = runtime::find_option(
options.options,
{"outetts.aligner_model_path", "outetts.forced_aligner_model_path"});
{"outetts.aligner_path", "outetts.aligner_model_path",
"outetts.forced_aligner_model_path"});
if (model_path.has_value()) {
return engine::models::qwen3_asr::load_qwen3_asr_assets(
std::filesystem::path(*model_path), "qwen3_forced_aligner");
Expand Down Expand Up @@ -485,7 +486,8 @@ OuteTTSSession::OuteTTSSession(runtime::TaskSpec task,
{"outetts.dac_weight_context_mb"},
1024ull * 1024ull * 1024ull),
runtime::parse_size_mb_option(options.options,
{"outetts.dac_graph_context_mb"},
{"outetts.dac_graph_arena_mb",
"outetts.dac_graph_context_mb"},
1536ull * 1024ull * 1024ull),
assets::TensorStorageType::F32),
mem_saver_(mem_saver_from_options(options)),
Expand Down Expand Up @@ -582,7 +584,7 @@ OuteTTSVoiceProfile OuteTTSSession::prepare_voice_profile(
throw std::runtime_error(
"OuteTTS voice cloning requires a GGUF with an embedded Qwen3 "
"Forced Aligner or --session-option "
"outetts.aligner_model_path=<path>");
"outetts.aligner_path=<path>");
}
aligner_session_ = std::make_unique<
engine::models::qwen3_forced_aligner::Qwen3ForcedAlignerSession>(
Expand Down
Loading