TTS OutieTTS 1.0 + cloning audio

CMakeLists.txt

@@ -257,6 +257,12 @@ add_library(engine_runtime STATIC
     src/models/omnivoice/assets.cpp
     src/models/omnivoice/tokenizer_text.cpp
     src/models/omnivoice/audio_tokenizer.cpp
+    src/models/outetts/assets.cpp
+    src/models/outetts/dac.cpp
+    src/models/outetts/llama.cpp
+    src/models/outetts/loader.cpp
+    src/models/outetts/session.cpp
+    src/models/outetts/tokenizer.cpp
     src/models/omnivoice/prompt_builder.cpp
     src/models/omnivoice/generator.cpp
     src/models/omnivoice/postprocess.cpp
@@ -716,6 +722,7 @@ if (ENGINE_BUILD_WARMBENCH)
     add_engine_warmbench(moss_tts_local_warm_bench tests/moss_tts_local/moss_tts_local_warm_bench.cpp)
     add_engine_warmbench(nemotron_asr_warm_bench tests/nemotron_asr/nemotron_asr_warm_bench.cpp)
     add_engine_warmbench(omnivoice_warm_bench tests/omnivoice/omnivoice_warm_bench.cpp)
+    add_engine_warmbench(outetts_warm_bench tests/outetts/outetts_warm_bench.cpp)
     add_engine_warmbench(pocket_tts_warm_bench tests/pocket_tts/pocket_tts_warm_bench.cpp)
     add_engine_warmbench(qwen3_asr_warm_bench tests/qwen3_asr/qwen3_asr_warm_bench.cpp)
     add_engine_warmbench(qwen3_forced_aligner_warm_bench tests/qwen3_forced_aligner/qwen3_forced_aligner_warm_bench.cpp)
@@ -785,6 +792,14 @@ if (ENGINE_BUILD_TESTS)
         COMMAND audio_chunking_test
     )
 
+    add_engine_unittest(outetts_generation_budget_test tests/unittests/test_outetts_generation_budget.cpp)
+    target_include_directories(outetts_generation_budget_test PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/tests/unittests)
+
+    add_test(
+        NAME outetts_generation_budget_test
+        COMMAND outetts_generation_budget_test
+    )
+
     add_engine_unittest(subtitle_formatter_test tests/unittests/test_subtitle_formatter.cpp)
 
     add_test(

README.md

@@ -60,6 +60,7 @@ audio.cpp would not be moving this quickly without generous contributors bringin
 | **miocodec** | audio codec, voice conversion backend | lang agnostic | MioCodec v2, 25 Hz, 44.1 kHz |
 | **miotts** | TTS, voice cloning | en, ja | MioTTS-1.7B |
 | **omnivoice** | TTS, voice cloning, voice design | 646+ langs | OmniVoice, Qwen3-0.6B based |
+| **outetts** | TTS, voice cloning | en, ar, zh, nl, fr, de, it, ja, ko, lt, ru, es, pt, be, bn, ka, hu, lv, fa, pl, sw, ta, uk | Llama-OuteTTS-1.0-1B |
 | **pocket_tts** | TTS, voice cloning | en, de, it, pt, es | PocketTTS-100M |
 | **nemotron_asr** | ASR | 100+ ASR prompt codes incl. auto | Nemotron 3.5 ASR Streaming 0.6B |
 | **qwen3_asr** | ASR | zh, en, yue, ar, de, fr, es, pt, id, it, ko, ru, th, vi, ja, tr, hi, ms, nl, sv, da, fi, pl, cs, fil, fa, el, ro, hu, mk | Qwen3-ASR-0.6B, Qwen3-ASR-1.7B-hf |
@@ -386,6 +387,7 @@ Recommended top-level install packages:
 | `moss_tts_local_v1_5` | MOSS-TTS-Local Transformer v1.5 | No |
 | `nemotron_asr` | Nemotron ASR | **Yes** |
 | `omnivoice` | OmniVoice | **Yes** |
+| `outetts_1_0_1b` | OuteTTS 1.0 1B with IBM DAC codec and Qwen3-aligned voice cloning | No |
 | `parakeet_tdt_0_6b_v3` | Parakeet TDT 0.6B v3 | **Yes** |
 | `pocket_tts` | PocketTTS | **Yes** |
 | `qwen3_asr_0_6b` | Qwen3 ASR 0.6B | **Yes** |
@@ -639,10 +641,10 @@ The framework also has a reusable GGUF tensor source and a streaming converter.
 container reader is shared by all model families; a family still has to list a `.gguf`
 checkpoint as one of its accepted assets because model configuration and tensor naming
 remain architecture-specific. Qwen3 ASR, Qwen3 Forced Aligner, Qwen3 TTS, Nemotron
-3.5 ASR, VibeVoice-ASR, Higgs Audio STT, Hviske ASR, and Citrinet ASR currently accept
+3.5 ASR, VibeVoice-ASR, Higgs Audio STT, Hviske ASR, Citrinet ASR, and OuteTTS currently accept
 `model.gguf` (including `speech_tokenizer/model.gguf` for TTS). The converter recursively embeds sidecar files
 up to 64 MiB by default using binary-safe metadata, including nested tokenizer models,
-and Qwen3 ASR, Nemotron ASR, VibeVoice-ASR, Higgs Audio STT, Hviske ASR, and Citrinet ASR
+and Qwen3 ASR, Nemotron ASR, VibeVoice-ASR, Higgs Audio STT, Hviske ASR, Citrinet ASR, and OuteTTS
 can load the resulting `model.gguf` as a standalone file. The converter embeds the selected
 package spec in new GGUF files. Standalone conversion with embedded sidecars is the default
 and fails if required package resources are missing. Pass `--no-sidecars` only to explicitly

docs/gguf.md

@@ -273,6 +273,7 @@ Status labels:
 | `moss_tts_nano` | Done | Pass | --- | Pass | No (similarity drift, frame drift, text large drift) |
 | `nemotron_asr` | Done | Pass | --- | Pass | Pass (minor filler drift) |
 | `omnivoice` | Done | Pass | --- | No (runtime assert, no audio) | No (runtime assert, no audio) |
+| `outetts` | Done | Pass (TTS + clone) | --- | --- | Pass (TTS + clone) |
 | `pocket_tts` | No | --- | --- | --- | --- |
 | `qwen3_asr` | Done | Pass | --- | Pass | Pass |
 | `qwen3_forced_aligner` | Done | Pass | --- | Pass | Pass |

docs/memory_saver.md

@@ -31,6 +31,7 @@ Native/default weights were used for all rows.
 - VoxCPM2 used the OpenAI-compatible offline speech endpoint with a 2048-character voice-design request, `seed=1234`, `max_tokens=512`, `num_inference_steps=10`, and `guidance_scale=2.0`. The default and `mem_saver` WAV outputs were byte-identical.
 - IndexTTS2 used a five-request server sequence with the same seeds and references for default and `mem_saver`: normal text, longer text, longer emotion-text request with a different reference, shorter text, then longer text.
 - Irodori TTS 500M used a five-request server sequence with the same seeds and options for default and `mem_saver`: reference text, longer reference text, longer no-reference emoji/style text, shorter reference text, then longer reference text.
+- OuteTTS 1.0 1B Q8 used a five-request long-lived-session sequence on an RTX 3090: repeated fixed-seed TTS, four-chunk long-form TTS, then repeated fixed-seed cloning with the same reference. CUDA timing is the mean of three fresh processes per mode, alternated to reduce ordering bias; memory saver was 0.35-0.53% slower per request and 0.44% slower over the mean sequence total. This is close to run-to-run variance and is not evidence of a speed benefit. The default and `mem_saver` WAV outputs were byte-identical. VRAM is total-device usage with no other CUDA workload; resident VRAM was sampled during a five-second post-sequence hold. Per-request timing and RTF are recorded in `docs/outetts_validation.md`.
 
 | Model | Mode | Peak VRAM | Resident VRAM | Server wall | Audio | RTF |
 |---|---|---:|---:|---:|---:|---:|
@@ -58,3 +59,5 @@ Native/default weights were used for all rows.
 | Irodori TTS 500M | mem_saver | 11222 MiB | 3570 MiB | 3276.203 ms | 95.4s | 0.0343418 |
 | Irodori TTS 500M 6000-char | default | 18693 MiB | 13367 MiB | 27185.7 ms | 777.92s | 0.0349466 |
 | Irodori TTS 500M 6000-char | mem_saver | 11588 MiB | 3609 MiB | 27828 ms | 777.92s | 0.0357724 |
+| OuteTTS 1.0 1B Q8 | default | 17653 MiB | 294 MiB | 29911.38 ms | 11.237s | 2.662 |
+| OuteTTS 1.0 1B Q8 | mem_saver | 5780 MiB | 294 MiB | 30043.25 ms | 11.237s | 2.674 |