feat: add native MiraTTS community model

CMakeLists.txt

@@ -1481,6 +1481,21 @@ audiocpp_add_model(soprano_tts
         engine::community_models::soprano_tts::make_soprano_tts_loader
 )
 
+audiocpp_add_model(mira_tts
+    SOURCES
+        src/community_models/mira_tts/assets.cpp
+        src/community_models/mira_tts/decoder.cpp
+        src/community_models/mira_tts/generator.cpp
+        src/community_models/mira_tts/processor.cpp
+        src/community_models/mira_tts/prompt.cpp
+        src/community_models/mira_tts/session.cpp
+        src/community_models/mira_tts/speaker_encoder.cpp
+    INCLUDES
+        engine/community_models/mira_tts/session.h
+    LOADERS
+        engine::community_models::mira_tts::make_mira_tts_loader
+)
+
 
 audiocpp_add_model(midashenglm_gen
     SOURCES
@@ -1910,6 +1925,7 @@ if (ENGINE_BUILD_WARMBENCH)
     add_engine_warmbench(irodori_tts_warm_bench tests/irodori_tts/irodori_tts_warm_bench.cpp)
     add_engine_warmbench(marblenet_vad_warm_bench tests/marblenet_vad/marblenet_vad_warm_bench.cpp)
     add_engine_warmbench(miocodec_warm_bench tests/miocodec/miocodec_warm_bench.cpp)
+    add_engine_warmbench(mira_tts_warm_bench tests/mira_tts/mira_tts_warm_bench.cpp)
     add_engine_warmbench(muscriptor_warm_bench tests/muscriptor/muscriptor_warm_bench.cpp)
     add_engine_warmbench(moss_tts_nano_warm_bench tests/moss_tts_nano/moss_tts_nano_warm_bench.cpp)
     add_engine_warmbench(moss_tts_local_warm_bench tests/moss_tts_local/moss_tts_local_warm_bench.cpp)

README.md

@@ -152,6 +152,7 @@ Community model ports live under `community_models` to make the ownership bounda
 | **kroko_asr** | ASR | de, en, es, fr, it, he, nl, pt, sv, tr | Safetensors, GGUF Q8 | Mirek [@mirek190](https://github.com/mirek190) | [Kroko Community ASR](docs/community_models/kroko_asr.md) native offline/streaming Zipformer2/RNN-T transcription with word timestamps |
 | **minimax_h3** | Video, Music, TTS/Dialogue | auto | GGUF Q4/INT8 | [@0xShug0](https://github.com/0xShug0) | [MiniMax-H3](docs/community_models/minimax_h3.md) text-to-audio/video generation with Q4_K and optional INT8 ConvRot DiT |
 | **minimax_music3** | Music | auto | GGUF Q4/Q8 | [@0xShug0](https://github.com/0xShug0), [@JoeMattie](https://github.com/JoeMattie) | [MiniMax Music 3](docs/community_models/minimax_music3.md) text-to-music generation with lyrics conditioning |
+| **mira_tts** | TTS, Clone | en | Local conversion | Mirek [@mirek190](https://github.com/mirek190) | [MiraTTS](docs/community_models/mira_tts.md) experimental native Qwen2 + ECAPA/Perceiver zero-shot voice cloning with progressive segment streaming (CC-BY-NC-SA-4.0 weights) |
 | **mms_forced_aligner** | Align | nl (nld), en (eng); pre-romanized Latin | Safetensors, GGUF 16/Q8 | Community | [MMS-300M-1130 Forced Aligner](docs/community_models/mms_forced_aligner.md) word-timestamp alignment from a wav2vec2 CTC checkpoint (safetensors or local GGUF) |
 | **moss_tts_local** | TTS, Clone, Ctrl | auto, optional language hint | GGUF | [@justinjohn0306](https://github.com/justinjohn0306) | MOSS-TTS-Local Transformer v1.5 support |
 | **moss_voicegen** | Voice Design | en, zh | GGUF | Joost [@jrohde](https://github.com/jrohde) | [MOSS-VoiceGenerator](docs/community_models/moss_voicegen.md) speech in a voice designed from a written instruction |

docs/community_models/mira_tts.md

@@ -0,0 +1,78 @@
+# MiraTTS
+
+MiraTTS is an experimental community port of
+[ysharma3501/MiraTTS](https://github.com/ysharma3501/MiraTTS), a zero-shot
+voice-cloning text-to-speech model. The native path includes the Qwen2 speech
+token generator, ECAPA-TDNN plus Perceiver reference encoder, finite-scalar
+speaker tokenizer, conditional acoustic processor, and DAC waveform decoder.
+
+## Model and license
+
+The upstream checkpoint is
+[YatharthS/MiraTTS](https://huggingface.co/YatharthS/MiraTTS). Its model card
+declares `CC-BY-NC-SA-4.0`; this is a non-commercial, attribution, share-alike
+license. Review that license before downloading or redistributing converted
+weights. audio.cpp does not redistribute the checkpoint.
+
+No ready-to-run GGUF package is published yet, so MiraTTS intentionally has no
+entry in the built-in download catalog. Convert a locally obtained upstream
+checkpoint with:
+
+```bash
+python tools/community_models/convert_mira_tts.py /path/to/MiraTTS /path/to/mira-native
+audiocpp_gguf \
+  --input language_model=/path/to/mira-native/language_model.safetensors \
+  --input speaker_encoder=/path/to/mira-native/speaker_encoder.safetensors \
+  --input processor=/path/to/mira-native/processor.safetensors \
+  --input decoder=/path/to/mira-native/decoder.safetensors \
+  --input upsampler=/path/to/mira-native/upsampler.safetensors \
+  --output /path/to/mira-native/mira-tts.gguf --type bf16 \
+  --family mira_tts --root /path/to/mira-native
+```
+
+Use BF16 for the first parity-oriented conversion. Converting the natively
+BF16 Qwen backbone to F16 can overflow and produce non-finite logits.
+
+The converter also imports the official FastBiCodec and FlashSR component
+checkpoints referenced by the upstream repository. Use `--help` to see its
+component path overrides.
+
+## Run
+
+MiraTTS requires reference audio. The CLI voice-cloning request accepts the
+converted model directory, target text, and a short clean reference WAV through
+the normal audio.cpp TTS/clone arguments. Sampling defaults reproduce upstream:
+temperature `0.8`, top-k `50`, top-p `0.95`, min-p `0.05`, and repetition
+penalty `1.2`.
+
+The upstream pipeline decodes at 16 kHz and applies its learned FlashSR
+upsampler. The native runtime executes both stages and returns 48 kHz audio.
+
+MiraTTS also exposes a streaming session. It splits long input at natural text
+boundaries, reuses one encoded speaker identity for the whole request, and emits
+each completed 48 kHz segment immediately. `text_chunk_size` controls the
+maximum segment size (160 codepoints by default), while `text_chunk_mode`
+selects the framework chunker. This is segment-level progressive synthesis;
+the acoustic processor, DAC, and FlashSR still decode each segment as a unit.
+
+The session caches one encoded reference voice by default, so repeated requests
+with the same audio do not rerun the speaker encoder. Increase the bounded cache
+with `--session-option reference_cache_slots=<n>`, or set it to `0` to disable
+reference reuse.
+
+## Validation status
+
+- Official checkpoint conversion: validated.
+- Native CUDA build: validated.
+- Native CUDA smoke synthesis through all converted model components: validated.
+- Segment-level streaming synthesis: validated through the native streaming
+  session and `/v1/audio/speech/live` route.
+- Deterministic upstream comparison with identical speech/context tokens:
+  validated (48 kHz waveform correlation 0.99996, SNR 41.1 dB).
+- End-to-end generation comparison: validated through matching tokenization,
+  identical 32-token speaker codes, and the first six greedy LM tokens. Later
+  autoregressive tokens can diverge between LMDeploy, Transformers, and the
+  native backend because of backend floating-point differences.
+
+Until end-to-end measurements are published, the family remains experimental
+and is not advertised as an installable WebUI package.

docs/community_models/models.md

@@ -26,6 +26,7 @@ Practical expectations:
 | **mms_forced_aligner** | Align | nl (nld), en (eng); pre-romanized Latin | Community | [MMS-300M-1130 Forced Aligner](mms_forced_aligner.md) word-timestamp alignment from a wav2vec2 CTC checkpoint (safetensors or local GGUF) |
 | **minimax_h3** | Video, Music, TTS/Dialogue | auto | [@0xShug0](https://github.com/0xShug0) | [MiniMax-H3](minimax_h3.md) text-to-audio/video generation with Q4_K and optional INT8 ConvRot DiT |
 | **minimax_music3** | Music | auto | [@0xShug0](https://github.com/0xShug0) | [MiniMax Music 3](minimax_music3.md) text-to-music generation with lyrics conditioning |
+| **mira_tts** | TTS, voice cloning | en | Mirek [@mirek190](https://github.com/mirek190) | [MiraTTS](mira_tts.md) experimental native Qwen2 + ECAPA/Perceiver zero-shot cloning; local conversion only (CC-BY-NC-SA-4.0 weights) |
 | **moss_tts_local** | TTS, voice cloning | auto, optional language hint | [@justinjohn0306](https://github.com/justinjohn0306) | [MOSS-TTS-Local Transformer v1.5](../models/moss_tts.md) support in the core model tree |
 | **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 | Mirek [@mirek190](https://github.com/mirek190) | [Llama-OuteTTS-1.0-1B](outetts.md) TTS and voice cloning support |
 | **voxcpm1** | TTS, voice cloning | zh, en, ja, ko | Community | [VoxCPM1](voxcpm1.md) tokenizer-free 0.5B TTS with 16 kHz output, streaming, and continuation-mode voice cloning |

docs/maintainers/model_specs.md

@@ -156,6 +156,11 @@ packages come from the same repo, put the shared source in
 `package_defaults.download` and keep package-level `download` only for
 overrides.
 
+Experimental ports may use an empty `packages` array while conversion and
+runtime validation are still local-only. In that case `ui.recommended_package`
+is omitted, so model managers do not advertise a download that cannot yet be
+loaded. Community and supported families must publish at least one package.
+
 ```json
 {
   "package_defaults": {

include/engine/community_models/mira_tts/assets.h

@@ -0,0 +1,56 @@
+#pragma once
+
+#include "engine/framework/assets/resource_bundle.h"
+#include "engine/framework/assets/tensor_source.h"
+
+#include <cstdint>
+#include <filesystem>
+#include <memory>
+
+namespace engine::community_models::mira_tts {
+
+struct MiraTTSConfig {
+    int64_t hidden_size = 896;
+    int64_t intermediate_size = 4864;
+    int64_t layers = 24;
+    int64_t attention_heads = 14;
+    int64_t kv_heads = 2;
+    int64_t head_dim = 64;
+    int64_t vocab_size = 166000;
+    int64_t max_position_embeddings = 32768;
+    float rms_norm_eps = 1.0e-6F;
+    float rope_theta = 1.0e6F;
+    int32_t bos_token_id = 151643;
+    int32_t eos_token_id = 151645;
+    int32_t speech_token_start = 155761;
+    int32_t speech_token_end = 163952;
+    int32_t prompt_speech_start = 165151;
+    int32_t sample_rate = 16000;
+    int32_t output_sample_rate = 48000;
+};
+
+struct MiraTTSAssets {
+    assets::ResourceBundle resources;
+    MiraTTSConfig config;
+    std::shared_ptr<const assets::TensorSource> language_model_weights;
+    std::shared_ptr<const assets::TensorSource> speaker_encoder_weights;
+    std::shared_ptr<const assets::TensorSource> processor_weights;
+    std::shared_ptr<const assets::TensorSource> decoder_weights;
+    std::shared_ptr<const assets::TensorSource> upsampler_weights;
+};
+
+struct MiraGenerationOptions {
+    int64_t max_new_tokens = 1024;
+    int64_t top_k = 50;
+    float top_p = 0.95F;
+    float min_p = 0.05F;
+    float temperature = 0.8F;
+    float repetition_penalty = 1.2F;
+    uint64_t seed = 0;
+    bool has_seed = false;
+};
+
+std::shared_ptr<const MiraTTSAssets> load_mira_tts_assets(
+    const std::filesystem::path & model_path);
+
+}  // namespace engine::community_models::mira_tts

include/engine/community_models/mira_tts/decoder.h

@@ -0,0 +1,33 @@
+#pragma once
+
+#include "engine/community_models/mira_tts/assets.h"
+#include "engine/framework/core/execution_context.h"
+#include "engine/framework/runtime/session.h"
+
+#include <cstddef>
+#include <cstdint>
+#include <memory>
+#include <vector>
+
+namespace engine::community_models::mira_tts {
+
+class MiraDecoder final {
+public:
+    MiraDecoder(
+        const MiraTTSAssets & assets,
+        core::ExecutionContext & execution,
+        size_t weight_context_bytes,
+        size_t graph_context_bytes,
+        assets::TensorStorageType storage_type);
+    ~MiraDecoder();
+
+    runtime::AudioBuffer decode(
+        const std::vector<float> & latents,
+        int64_t frames);
+
+private:
+    struct Impl;
+    std::unique_ptr<Impl> impl_;
+};
+
+}  // namespace engine::community_models::mira_tts

include/engine/community_models/mira_tts/generator.h

@@ -0,0 +1,34 @@
+#pragma once
+
+#include "engine/community_models/mira_tts/assets.h"
+#include "engine/framework/core/execution_context.h"
+
+#include <cstddef>
+#include <cstdint>
+#include <memory>
+#include <vector>
+
+namespace engine::community_models::mira_tts {
+
+class MiraGenerator final {
+public:
+    MiraGenerator(
+        const MiraTTSAssets & assets,
+        core::ExecutionContext & execution,
+        size_t prefill_graph_arena_bytes,
+        size_t decode_graph_arena_bytes,
+        size_t weight_context_bytes,
+        assets::TensorStorageType weight_storage_type);
+    ~MiraGenerator();
+
+    std::vector<int32_t> generate(
+        const std::vector<int32_t> & prompt_ids,
+        const MiraGenerationOptions & options);
+    void release_runtime_graphs();
+
+private:
+    struct Impl;
+    std::unique_ptr<Impl> impl_;
+};
+
+}  // namespace engine::community_models::mira_tts

include/engine/community_models/mira_tts/processor.h

@@ -0,0 +1,33 @@
+#pragma once
+
+#include "engine/community_models/mira_tts/assets.h"
+#include "engine/framework/core/execution_context.h"
+
+#include <cstddef>
+#include <cstdint>
+#include <memory>
+#include <vector>
+
+namespace engine::community_models::mira_tts {
+
+class MiraAcousticProcessor final {
+public:
+    MiraAcousticProcessor(
+        const MiraTTSAssets & assets,
+        core::ExecutionContext & execution,
+        size_t weight_context_bytes,
+        size_t graph_context_bytes,
+        assets::TensorStorageType linear_storage_type,
+        assets::TensorStorageType conv_storage_type);
+    ~MiraAcousticProcessor();
+
+    std::vector<float> process(
+        const std::vector<int32_t> & speech_codes,
+        const std::vector<int32_t> & context_codes);
+
+private:
+    struct Impl;
+    std::unique_ptr<Impl> impl_;
+};
+
+}  // namespace engine::community_models::mira_tts

include/engine/community_models/mira_tts/prompt.h

@@ -0,0 +1,26 @@
+#pragma once
+
+#include "engine/community_models/mira_tts/assets.h"
+
+#include <cstdint>
+#include <memory>
+#include <string>
+#include <vector>
+
+namespace engine::community_models::mira_tts {
+
+class MiraPromptBuilder final {
+public:
+    explicit MiraPromptBuilder(std::shared_ptr<const MiraTTSAssets> assets);
+    ~MiraPromptBuilder();
+
+    std::vector<int32_t> build(
+        const std::string & text,
+        const std::vector<int32_t> & context_codes) const;
+
+private:
+    struct Impl;
+    std::unique_ptr<Impl> impl_;
+};
+
+}  // namespace engine::community_models::mira_tts