Skip to content

libelizainference ABI v17: Paradee + misaki TTS entries, ASR cancel/supported, TTS cancel and evict for Kokoro/Paradee #59

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
lalalune merged 2 commits into from
Oct 14, 2026
Merged

libelizainference ABI v17: Paradee + misaki TTS entries, ASR cancel/supported, TTS cancel and evict for Kokoro/Paradee #59

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
8e7ced9
feat(omnivoice): libelizainference ABI v17 (Paradee, ASR cancel, TTS …
lalalune Oct 14, 2026
76ed981
test(omnivoice): run the v17 FFI tests in CI and fix racy check counters
lalalune Oct 14, 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
84 changes: 84 additions & 0 deletions tools/omnivoice/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -232,7 +232,28 @@ endif()
# (the fused build links mtmd anyway), opt out with -DELIZA_ENABLE_VISION=OFF.
option(ELIZA_ENABLE_VISION "Build the fused mmproj vision-describe ABI (v9)" ON)

# Paradee-8M TTS + native misaki G2P (ABI v17). Folded in only when BOTH
# paradee_lib and misaki_g2p exist (tools/paradee, tools/misaki-g2p; they must
# be added before this directory, like kokoro_lib). PRIVATE: the
# eliza_paradee:: / eliza_misaki:: C++ APIs stay behind the
# eliza_inference_paradee_* surface.
function(eliza_inference_fold_paradee target)
if(TARGET paradee_lib AND TARGET misaki_g2p)
target_link_libraries(${target} PRIVATE paradee_lib misaki_g2p)
target_include_directories(${target} PRIVATE
${CMAKE_CURRENT_SOURCE_DIR}/../paradee/include
${CMAKE_CURRENT_SOURCE_DIR}/../misaki-g2p/include)
target_compile_definitions(${target} PRIVATE ELIZA_ENABLE_PARADEE)
endif()
endfunction()

if(TARGET mtmd)
if(TARGET paradee_lib AND TARGET misaki_g2p)
message(STATUS "elizainference: Paradee TTS + misaki G2P enabled (ABI v17)")
elseif(TARGET paradee_lib OR TARGET misaki_g2p)
message(WARNING "elizainference: Paradee needs both paradee_lib and misaki_g2p; "
"building without Paradee (paradee_supported() == 0)")
endif()
add_library(elizainference SHARED
${OMNIVOICE_CORE_SOURCES}
${OMNIVOICE_FFI_SOURCES})
Expand Down Expand Up @@ -283,6 +304,7 @@ if(TARGET mtmd)
${CMAKE_CURRENT_SOURCE_DIR}/../kokoro/include)
target_compile_definitions(elizainference PRIVATE ELIZA_ENABLE_KOKORO)
endif()
eliza_inference_fold_paradee(elizainference)
set_target_properties(elizainference PROPERTIES
OUTPUT_NAME elizainference
POSITION_INDEPENDENT_CODE ON
Expand Down Expand Up @@ -337,6 +359,7 @@ if(TARGET mtmd)
${CMAKE_CURRENT_SOURCE_DIR}/../kokoro/include)
target_compile_definitions(elizainference_static PRIVATE ELIZA_ENABLE_KOKORO)
endif()
eliza_inference_fold_paradee(elizainference_static)
set_target_properties(elizainference_static PROPERTIES
OUTPUT_NAME elizainference
POSITION_INDEPENDENT_CODE ON)
Expand All @@ -345,6 +368,67 @@ if(TARGET mtmd)
set_target_properties(elizainference_static PROPERTIES OUTPUT_NAME elizainference_static)
endif()
add_dependencies(elizainference_static omnivoice-version)

# FFI tests (ABI v17).
# eliza-inference-test-abi links the real libelizainference: ABI
# string, probes, cancel/evict no-ops, and
# (opt-in, ELIZA_TEST_ASR_DIR) a real-model
# ASR cancel during prefill.
# eliza-inference-test-paradee-asr compiles the FFI bridge into the test
# (ELIZA_INFERENCE_TEST_HOOKS) against the
# scripted Paradee/misaki doubles in
# tests/paradee-contract/: load/synth/cancel
# race, chunking, evict -> reload, ASR cancel
# during prefill and decode.
option(ELIZA_INFERENCE_BUILD_FFI_TESTS "Build the libelizainference FFI tests" ${LLAMA_BUILD_TESTS})
if(ELIZA_INFERENCE_BUILD_FFI_TESTS)
if(NOT WIN32) # POSIX threads / mkdtemp
add_executable(eliza-inference-test-abi tests/ffi-abi.c)
set_target_properties(eliza-inference-test-abi PROPERTIES
C_STANDARD 11
C_STANDARD_REQUIRED ON)
target_include_directories(eliza-inference-test-abi PRIVATE
${CMAKE_CURRENT_SOURCE_DIR}/include)
find_package(Threads REQUIRED)
target_link_libraries(eliza-inference-test-abi PRIVATE elizainference Threads::Threads)
if(TARGET paradee_lib AND TARGET misaki_g2p)
target_compile_definitions(eliza-inference-test-abi PRIVATE ELIZA_TEST_EXPECT_PARADEE=1)
endif()
if(BUILD_TESTING)
add_test(NAME eliza-inference-test-abi COMMAND eliza-inference-test-abi)
# "main" is the label CI runs (ctest -L main); unlabelled
# tests are built but never executed there.
set_property(TEST eliza-inference-test-abi PROPERTY LABELS main)
endif()
endif()

add_executable(eliza-inference-test-paradee-asr
tests/ffi-paradee-asr.cpp
tests/paradee-contract/fake-paradee.cpp)
target_compile_features(eliza-inference-test-paradee-asr PRIVATE cxx_std_17)
target_compile_definitions(eliza-inference-test-paradee-asr PRIVATE
${OMNIVOICE_COMPILE_DEFINITIONS}
ELIZA_ENABLE_PARADEE ELIZA_INFERENCE_TEST_HOOKS)
# The contract doubles come first so "paradee.h" / "misaki-g2p.h"
# resolve to them even when the real libraries are also configured.
target_include_directories(eliza-inference-test-paradee-asr BEFORE PRIVATE
${CMAKE_CURRENT_SOURCE_DIR}/tests/paradee-contract)
target_include_directories(eliza-inference-test-paradee-asr PRIVATE
${CMAKE_CURRENT_SOURCE_DIR}/src
${CMAKE_CURRENT_SOURCE_DIR}/include
${CMAKE_CURRENT_BINARY_DIR}
${OMNIVOICE_LLAMA_PRIVATE_INCLUDE_DIRS}
${CMAKE_CURRENT_SOURCE_DIR}/../mtmd
${CMAKE_CURRENT_SOURCE_DIR}/../../include
${CMAKE_CURRENT_SOURCE_DIR}/../../common)
target_link_libraries(eliza-inference-test-paradee-asr PRIVATE
omnivoice_lib llama mtmd eliza_voice_classifiers llama-common)

if(BUILD_TESTING)
add_test(NAME eliza-inference-test-paradee-asr COMMAND eliza-inference-test-paradee-asr)
set_property(TEST eliza-inference-test-paradee-asr PROPERTY LABELS main)
endif()
endif()
endif()

# llama-server wiring — mount the OmniVoice POST /v1/audio/speech route
Expand Down
195 changes: 188 additions & 7 deletions tools/omnivoice/include/eliza-inference-ffi.h
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
/*
* libelizainference FFI ABI v16.
* libelizainference FFI ABI v17.
*
* (Banner tracks ELIZA_INFERENCE_ABI_VERSION below; the per-version history is
* at the end of this header preamble, newest first.)
Expand Down Expand Up @@ -134,6 +134,34 @@ extern "C" {
* load and refuses to bind if they disagree.
*
* Changelog:
* v17: Paradee-8M TTS with native misaki G2P, plus cancellation and
* eviction for every TTS head and for ASR. All additions are new
* symbols; no existing signature changes.
* - `eliza_inference_paradee_supported/load/sample_rate/g2p_kind/
* phonemize/synthesize_alloc/synthesize_phonemes_alloc`. Built
* with `ELIZA_ENABLE_PARADEE` (TARGETs paradee_lib + misaki_g2p).
* Text goes through the native misaki lexicon, is split into
* <= 510-id chunks, and the PCM is concatenated. PCM is released
* with `eliza_inference_free_pcm`.
* - `eliza_inference_cancel_tts` now also stops in-flight Kokoro and
* Paradee synthesis (ELIZA_ERR_CANCELLED). Paradee polls the flag
* inside its graph and between chunks; Kokoro checks it before and
* after the forward pass (kokoro_lib has no in-graph hook).
* - `eliza_inference_asr_cancel` + `eliza_inference_asr_supported`.
* Cancel is an atomic flag (no asr_mutex), checked by the llama
* abort callback during prefill/decode, between mtmd chunks, and
* before every decode token.
* - `mmap_evict("tts")` frees Kokoro and Paradee (+ lexicon) as well
* as OmniVoice, keeping their paths; the next synth or
* `mmap_acquire("tts")` reloads them. `mmap_acquire("tts")` on a
* bundle without OmniVoice weights succeeds when a Kokoro or
* Paradee head is configured on the ctx.
* - The OmniVoice GGUF picker skips `tts/paradee/` and `paradee*`
* files, as it already skipped `tts/kokoro/`.
* Streaming ASR (`asr_stream_*`) is unchanged and still reports
* `asr_stream_supported() == 0`.
* v16: context-local text-model GPU selection
* (`eliza_inference_create_with_options`).
* v15: Kokoro synthesis can return an exact-size, library-owned PCM buffer
* through `eliza_inference_kokoro_synthesize_alloc` and
* `eliza_inference_kokoro_synthesize_ipa_alloc`; callers release it with
Expand Down Expand Up @@ -226,9 +254,9 @@ extern "C" {
* v7: real Silero VAD (same symbol surface as v6).
* v6: fused wake-word, speaker, diarizer.
*/
#define ELIZA_INFERENCE_ABI_VERSION 16
#define ELIZA_INFERENCE_ABI_VERSION 17

/* Returns a static, NUL-terminated string of the form "16" matching
/* Returns a static, NUL-terminated string of the form "17" matching
* ELIZA_INFERENCE_ABI_VERSION at the time the library was built. The
* pointer is owned by the library — do NOT free. */
const char * eliza_inference_abi_version(void);
Expand All @@ -244,7 +272,34 @@ const char * eliza_inference_abi_version(void);
#define ELIZA_ERR_FFI_FAULT -4 /* mmap/madvise/syscall failure */
#define ELIZA_ERR_OOM -5 /* allocation failure */
#define ELIZA_ERR_ABI_MISMATCH -6 /* loader vs library disagree */
#define ELIZA_ERR_CANCELLED -7 /* caller requested cancellation (chunk cb / cancel_tts) */
#define ELIZA_ERR_CANCELLED -7 /* caller requested cancellation (chunk cb / cancel_tts / asr_cancel) */

/* ---- Capability probes (per capability) --------------------------- *
*
* Probe before binding a surface; an absent symbol (older library) means
* unsupported. All probes are build properties and take no ctx unless noted.
*
* capability probe since
* ---------------------- -------------------------------------- -----
* batch ASR + cancel eliza_inference_asr_supported v17
* ASR word timestamps eliza_inference_asr_timestamps_supported v12
* streaming ASR eliza_inference_asr_stream_supported v2 (0 today)
* OmniVoice stream+cancel eliza_inference_tts_stream_supported v2
* Kokoro TTS eliza_inference_kokoro_supported v10
* Kokoro G2P kind eliza_inference_kokoro_g2p_kind(ctx) v14
* Paradee TTS + misaki eliza_inference_paradee_supported v17
* misaki fallback kind eliza_inference_paradee_g2p_kind(ctx) v17
* Silero VAD eliza_inference_vad_supported v3/v7
* wake word eliza_inference_wakeword_supported v5
* speaker embedding eliza_inference_speaker_supported v6
* diarizer eliza_inference_diariz_supported v6
* streaming LLM eliza_inference_llm_stream_supported v4
* MTP / KV-cache quant eliza_inference_llm_{mtp,kv_quant}_supported v8
* embeddings eliza_inference_embed_supported v9
* vision describe eliza_inference_vision_supported v9
* tokenize / detokenize eliza_inference_tokenize_supported v9
* end-of-turn scoring eliza_inference_llm_eot_supported v11
*/

/* ---- Lifecycle ------------------------------------------------------ */

Expand Down Expand Up @@ -279,7 +334,10 @@ void eliza_inference_destroy(EliInferenceContext * ctx);
* by calling `mmap_evict(...)` on the same region names.
*
* `region_name` is a stable string in the set:
* - "tts" : OmniVoice weights (mmap of tts/omnivoice-*.gguf)
* - "tts" : every TTS head on the ctx. OmniVoice weights (mmap of
* tts/omnivoice-*.gguf), plus the Kokoro model + voice and the
* Paradee model + misaki lexicon once they have been loaded with
* `kokoro_load` / `paradee_load`.
* - "asr" : ASR weights (mmap of asr/...)
* - "text" : text+vision weights (kept hot — always acquired)
* - "dflash" : drafter weights (kept hot — always acquired)
Expand All @@ -291,7 +349,20 @@ void eliza_inference_destroy(EliInferenceContext * ctx);
* voice-only region to minimize voice-off RSS. Callers must treat an
* evicted region as unavailable until a later `mmap_acquire(region)`.
* The "text" and "dflash" regions are allowed to be no-ops because the
* text runtime keeps them hot for voice-off text turns. */
* text runtime keeps them hot for voice-off text turns.
*
* Region semantics (ABI v17):
* evict("tts") frees OmniVoice, Kokoro and Paradee + lexicon. Each head
* is freed under its own lock, so a synth in flight on that
* head finishes first. Kokoro / Paradee paths are kept.
* acquire("tts") reloads any Kokoro / Paradee head evicted above, then
* loads OmniVoice. It succeeds without OmniVoice weights in
* the bundle when a Kokoro or Paradee head is configured.
* Kokoro / Paradee synth calls also reload an evicted head themselves, so
* acquire("tts") is optional for them. OmniVoice synth still requires it.
* evict("asr") frees the ASR model, context, mmproj and sampler
* (unchanged); acquire("asr") reloads them.
* "text", "dflash", "vad": no-ops. */
int eliza_inference_mmap_acquire(
EliInferenceContext * ctx,
const char * region_name,
Expand Down Expand Up @@ -373,7 +444,15 @@ int eliza_inference_tts_synthesize_stream(
* `eliza_inference_tts_synthesize_stream` on another thread). The
* in-flight call returns ELIZA_ERR_CANCELLED at the next kernel
* boundary. Returns ELIZA_OK whether or not a forward pass was running
* (cancelling nothing is not an error). */
* (cancelling nothing is not an error).
*
* ABI v17: also cancels in-flight Kokoro (`kokoro_synthesize*`) and Paradee
* (`paradee_synthesize*`) calls. Each engine has its own flag and clears it
* when a call on that engine starts, so a cancel never carries over into a
* call that starts after it. Paradee stops inside the graph or between
* chunks; Kokoro is checked before and after its forward pass, so a Kokoro
* call finishes its current pass and then returns ELIZA_ERR_CANCELLED with
* no audio. */
int eliza_inference_cancel_tts(
EliInferenceContext * ctx,
char ** out_error);
Expand Down Expand Up @@ -483,6 +562,87 @@ int eliza_inference_kokoro_synthesize_ipa_alloc(

void eliza_inference_free_pcm(float * pcm);

/* ---- Paradee-8M TTS + native misaki G2P (ABI v17) ----------------- *
*
* Paradee-8M v1.0 (Apache-2.0, distilled from Kokoro-82M): one voice
* (af_heart), 24 kHz mono fp32 out. The model input is misaki phoneme ids
* (the Kokoro vocab) with a 0 pad at each end, at most 512 ids, plus a speed
* scalar. G2P runs natively through the misaki English lexicon + fallback
* linked into this library; no ONNX, no host phonemizer.
*
* Build flag `ELIZA_ENABLE_PARADEE` (TARGETs paradee_lib + misaki_g2p). In a
* build without it `paradee_supported() == 0` and every other entry returns
* ELIZA_ERR_NOT_IMPLEMENTED. Model + lexicon are owned by the ctx, guarded by
* their own lock (independent of OmniVoice / Kokoro), freed by
* `mmap_evict("tts")` (reloaded lazily) and by `eliza_inference_destroy`.
* `eliza_inference_cancel_tts` cancels an in-flight Paradee synth. */

/* 1 when this build linked paradee_lib + misaki_g2p; 0 otherwise. */
int eliza_inference_paradee_supported(void);

/* Load the Paradee GGUF and the misaki lexicon into `ctx`. A repeat call with
* the same paths is a no-op. `espeak_data_dir` may be NULL: the lexicon then
* uses its rules fallback for out-of-vocabulary words (see `g2p_kind`).
* On failure the previously loaded model, if any, stays loaded. Returns
* ELIZA_OK or a negative ELIZA_* code with `*out_error` set. */
int eliza_inference_paradee_load(
EliInferenceContext * ctx,
const char * gguf_path,
const char * lexicon_path,
const char * espeak_data_dir,
char ** out_error);

/* Sample rate of the loaded model (24000 for v1.0). Still answers after
* `mmap_evict("tts")`. Negative ELIZA_* if `paradee_load` never succeeded. */
int eliza_inference_paradee_sample_rate(EliInferenceContext * ctx);

/* How the loaded misaki lexicon phonemizes words it does not contain. */
#define ELIZA_PARADEE_G2P_LEXICON_ONLY 0 /* no fallback */
#define ELIZA_PARADEE_G2P_RULES 1 /* native letter-to-sound rules */
#define ELIZA_PARADEE_G2P_ESPEAK 2 /* eSpeak NG, converted to misaki spelling */
/* Returns one of ELIZA_PARADEE_G2P_*, or a negative ELIZA_* code when no
* lexicon was loaded (or on a non-Paradee build). */
int eliza_inference_paradee_g2p_kind(EliInferenceContext * ctx);

/* Run the native G2P only: UTF-8 English text in, misaki phoneme string out.
* On success `*out_utf8` is a NUL-terminated heap string (free it with
* `eliza_inference_free_string`) and the return value is its byte length. */
int eliza_inference_paradee_phonemize(
EliInferenceContext * ctx,
const char * text,
size_t text_len,
char ** out_utf8,
char ** out_error);

/* Text -> PCM. Native misaki G2P, then the phoneme string is split into
* <= 510-id chunks (sentence ends, then clause punctuation, then spaces),
* each chunk is padded and synthesized, and the PCM is concatenated. `speed`
* scales the speaking rate (1.0 = native; <= 0 or non-finite means 1.0).
* On success `*out_pcm` is a malloc-owned buffer of `*out_samples` floats at
* `paradee_sample_rate` (NULL only when `*out_samples == 0`); release it with
* `eliza_inference_free_pcm`. Returns ELIZA_OK, ELIZA_ERR_CANCELLED after
* `eliza_inference_cancel_tts`, or another negative ELIZA_* code. */
int eliza_inference_paradee_synthesize_alloc(
EliInferenceContext * ctx,
const char * text,
size_t text_len,
float speed,
float ** out_pcm,
size_t * out_samples,
char ** out_error);

/* Same as `paradee_synthesize_alloc`, but the input is a misaki phoneme
* string (as returned by `paradee_phonemize`); G2P is skipped. Characters
* outside the model vocab are dropped. */
int eliza_inference_paradee_synthesize_phonemes_alloc(
EliInferenceContext * ctx,
const char * misaki_phonemes,
size_t phonemes_len,
float speed,
float ** out_pcm,
size_t * out_samples,
char ** out_error);

/* ---- OmniVoice reference encode (ABI v4) -------------------------- *
*
* Encode a 24 kHz mono fp32 PCM buffer through the OmniVoice tokenizer
Expand Down Expand Up @@ -778,6 +938,27 @@ int eliza_inference_asr_transcribe(
size_t max_text_bytes,
char ** out_error);

/* ---- ASR probe + cancel (ABI v17) --------------------------------- */

/* 1 when this build has the batch ASR path (mtmd audio encoder compiled in).
* Always 1 for libelizainference, which is only built with mtmd. Whether the
* bundle has ASR weights is answered by `mmap_acquire("asr")`. */
int eliza_inference_asr_supported(void);

/* Cancel the `asr_transcribe` / `asr_transcribe_timed` call in flight on
* `ctx` (from another thread). Sets an atomic flag without taking the ASR
* lock. The flag is checked by the llama abort callback during the prefill
* and decode graphs, between prompt chunks, and before every decode token;
* the audio encoder pass itself has no abort hook, so a cancel during it
* lands when that pass ends. The cancelled call returns ELIZA_ERR_CANCELLED.
* Each transcription clears the flag when it starts (after it holds the ASR
* lock), so a cancel with nothing in flight has no effect on later calls.
* Returns ELIZA_OK (cancelling nothing is not an error), or
* ELIZA_ERR_INVALID_ARG for a NULL ctx. */
int eliza_inference_asr_cancel(
EliInferenceContext * ctx,
char ** out_error);

/* ---- ASR word timestamps (ABI v12) -------------------------------- */

/* Capability probe: returns 1 when this build can emit per-word timestamps
Expand Down
Loading
Loading