openvinotoolkit · michalkulakowski · Jul 3, 2026 · Jul 6, 2026 · Jul 6, 2026 · Jul 6, 2026
diff --git a/Dockerfile.redhat b/Dockerfile.redhat
@@ -315,6 +315,12 @@ RUN rm -f /usr/lib64/cmake/OpenSSL/OpenSSLConfig.cmake
 # hadolint ignore=SC2046
 RUN bazel build --jobs=$JOBS ${debug_bazel_flags} ${minitrace_flags} //src:ovms $(if [ "$OPTIMIZE_BUILDING_TESTS" == "1" ] ; then echo -n //src:ovms_test; fi)
 
+# espeak-ng is built as a separate step, independent of the OVMS binary.
+# Set ESPEAK=0 to skip the espeak build.
+ARG ESPEAK=1
+# hadolint ignore=DL3059
+RUN if [ "$ESPEAK" == "1" ]; then bazel build --jobs=$JOBS ${debug_bazel_flags} //third_party:espeak_ng //third_party:espeak_ng_data; fi
+
 # Tests execution
 COPY ci/check_coverage.bat /ovms/
 ARG CHECK_COVERAGE=0

diff --git a/Dockerfile.ubuntu b/Dockerfile.ubuntu
@@ -331,6 +331,12 @@ ARG OPTIMIZE_BUILDING_TESTS=0
 # hadolint ignore=SC2046
 RUN if [ "$FUZZER_BUILD" == "0" ]; then bazel build --jobs=$JOBS ${debug_bazel_flags} ${minitrace_flags} //src:ovms $(if [ "${OPTIMIZE_BUILDING_TESTS}" == "1" ] ; then echo -n //src:ovms_test; fi); fi;
 
+# espeak-ng is built as a separate step, independent of the OVMS binary.
+# Set ESPEAK=0 to skip the espeak build.
+ARG ESPEAK=1
+# hadolint ignore=DL3059
+RUN if [ "$ESPEAK" == "1" ]; then bazel build --jobs=$JOBS ${debug_bazel_flags} //third_party:espeak_ng //third_party:espeak_ng_data; fi
+
 ARG RUN_TESTS=0
 RUN if [ "$RUN_TESTS" == "1" ] ; then mkdir -p demos/common/export_models/ && mv export_model.py demos/common/export_models/ && ./prepare_llm_models.sh /ovms/src/test/llm_testing docker && ./run_unit_tests.sh ; fi
 

diff --git a/Makefile b/Makefile
@@ -148,13 +148,8 @@ else ifeq ($(findstring redhat,$(BASE_OS)),redhat)
 else
   $(error BASE_OS must be either ubuntu or redhat)
 endif
-ifeq ($(ESPEAK),1)
-  ESPEAK_PARAMS = " --//:espeak=on"
-else
-  ESPEAK_PARAMS = " --//:espeak=off"
-endif
-CAPI_FLAGS = "--strip=$(STRIP)"$(BAZEL_DEBUG_BUILD_FLAGS)"  --config=mp_off_py_off"$(OV_TRACING_PARAMS)$(TARGET_DISTRO_PARAMS)$(ESPEAK_PARAMS)
-BAZEL_DEBUG_FLAGS="--strip=$(STRIP)"$(BAZEL_DEBUG_BUILD_FLAGS)$(DISABLE_PARAMS)$(FUZZER_BUILD_PARAMS)$(OV_TRACING_PARAMS)$(TARGET_DISTRO_PARAMS)$(ESPEAK_PARAMS)$(REPO_ENV)
+CAPI_FLAGS = "--strip=$(STRIP)"$(BAZEL_DEBUG_BUILD_FLAGS)"  --config=mp_off_py_off"$(OV_TRACING_PARAMS)$(TARGET_DISTRO_PARAMS)
+BAZEL_DEBUG_FLAGS="--strip=$(STRIP)"$(BAZEL_DEBUG_BUILD_FLAGS)$(DISABLE_PARAMS)$(FUZZER_BUILD_PARAMS)$(OV_TRACING_PARAMS)$(TARGET_DISTRO_PARAMS)$(REPO_ENV)
 
 # Option to Override release image.
 # Release image OS *must have* glibc version >= glibc version on BASE_OS:
@@ -249,7 +244,8 @@ BUILD_ARGS = --build-arg http_proxy=$(HTTP_PROXY)\
 	--build-arg JOBS=$(JOBS)\
 	--build-arg CAPI_FLAGS=$(CAPI_FLAGS)\
 	--build-arg VERBOSE_LOGS=$(VERBOSE_LOGS)\
-	--build-arg KONFLUX=$(KONFLUX)
+	--build-arg KONFLUX=$(KONFLUX)\
+	--build-arg ESPEAK=$(ESPEAK)
 
 
 .PHONY: default docker_build \

diff --git a/common_settings.bzl b/common_settings.bzl
@@ -20,7 +20,7 @@
 load("@bazel_skylib//lib:selects.bzl", "selects")
 load("@mediapipe//mediapipe/framework:more_selects.bzl", "more_selects")
 load("@bazel_skylib//rules:common_settings.bzl", "string_flag")
-load("//:distro.bzl", "distro_flag", "espeak_flag")
+load("//:distro.bzl", "distro_flag")
 
 # cc_library rule wrapper that will accept the same arguments but if user will not provide
 # copts, linkopts, local_defines it will set them to the defaults
@@ -58,7 +58,6 @@ def ovms_cc_library(**kwargs):
 
 def create_config_settings():
     distro_flag()
-    espeak_flag()
     native.config_setting(
         name = "disable_mediapipe",
         define_values = {

diff --git a/create_package.sh b/create_package.sh
@@ -36,6 +36,7 @@ ESPEAK_DATA_SRC=$(find /ovms/bazel-out/k8-*/bin/external/espeak_ng -type d -name
 if [ -n "$ESPEAK_DATA_SRC" ] && [ -d "$ESPEAK_DATA_SRC" ] ; then
     mkdir -p /ovms_release/share
     cp -rL "$ESPEAK_DATA_SRC" /ovms_release/share/ ;
+    chmod -R u+w /ovms_release/share/espeak-ng-data 2>/dev/null || true ;
 fi
 # Resolve the packaged eSpeak shared object dynamically so version bumps
 # do not require touching this script.

diff --git a/demos/audio/README.md b/demos/audio/README.md
@@ -8,116 +8,80 @@ Check supported [Speech Recognition Models](https://openvinotoolkit.github.io/op
 
 ## Prerequisites
 
-**OVMS version 2025.4** This demo require version 2025.4 or nightly release.
-
-**Model preparation**: Python 3.10 or higher with pip
+**OVMS version 2026.3** This demo require version 2026.3 or nightly release.
 
 **Model Server deployment**: Installed Docker Engine or OVMS binary package according to the [baremetal deployment guide](../../docs/deploying_server_baremetal.md)
 
 **Client**: curl or Python for using OpenAI client package
 
 ## Speech generation
-### Prepare speaker embeddings
-When generating speech you can use default speaker voice or you can prepare your own speaker embedding file. Here you can see how to do it with downloaded file from online repository, but you can try with your own speech recording as well:
-```console
-pip install -r https://raw.githubusercontent.com/openvinotoolkit/model_server/refs/heads/main/demos/audio/requirements.txt
-mkdir audio_samples
-curl --create-dirs "https://www.voiptroubleshooter.com/open_speech/american/OSR_us_000_0032_8k.wav" -o audio_samples/audio.wav
-curl --create-dirs https://raw.githubusercontent.com/openvinotoolkit/model_server/refs/heads/main/demos/audio/create_speaker_embedding.py -o models/speakers/create_speaker_embedding.py
-python models/speakers/create_speaker_embedding.py audio_samples/audio.wav models/speakers/voice1.bin
-```
-
+Kokoro is the primary example in this demo, but SpeechT5 remains supported for existing deployments.
 ### Model preparation
-Supported models should use the topology of [microsoft/speecht5_tts](https://huggingface.co/microsoft/speecht5_tts) which needs to be converted to IR format before using in OVMS.
-
-Specific OVMS pull mode example for models requiring conversion is described in the [Ovms pull mode](../../docs/pull_hf_models.md#pulling-models-outside-openvino-organization)
-
-Or you can use the python export_model.py script described below.
-
-Here, the original Text to Speech model will be converted to IR format and optionally quantized.
-That ensures faster initialization time, better performance and lower memory consumption.
-Execution parameters will be defined inside the `graph.pbtxt` file.
-
-Download export script, install it's dependencies and create directory for the models:
-```console
-curl https://raw.githubusercontent.com/openvinotoolkit/model_server/refs/heads/main/demos/common/export_models/export_model.py -o export_model.py
-pip install -r https://raw.githubusercontent.com/openvinotoolkit/model_server/refs/heads/main/demos/common/export_models/requirements.txt
-mkdir models
-```
-
-Run `export_model.py` script to download and quantize the model:
-
-> **Note:** The users in China need to set environment variable HF_ENDPOINT="https://hf-mirror.com" before running the export script to connect to the HF Hub.
-> **Note:** Exporting `microsoft/speecht5_tts` model requires Python 3.10
-
-**CPU**
-```console
-python export_model.py text2speech --source_model microsoft/speecht5_tts --weight-format fp16 --model_name microsoft/speecht5_tts --config_file_path models/config.json --model_repository_path models --overwrite_models --vocoder microsoft/speecht5_hifigan --speaker_name voice1 --speaker_path models/speakers/voice1.bin
-```
+This demo uses a pre-exported OpenVINO IR model [luis-castillo/Kokoro-82M-OpenVINO-FP16-OVMS](https://huggingface.co/luis-castillo/Kokoro-82M-OpenVINO-FP16-OVMS) available on HuggingFace.
+The model can be pulled directly by OVMS without any conversion step.
 
-> **Note:** Change the `--weight-format` to quantize the model to `int8` precision to reduce memory consumption and improve performance.
-> **Note:** `speaker_name` and `speaker_path` may be omitted if the default model voice is sufficient
+> **Note:** The users in China need to set environment variable HF_ENDPOINT="https://hf-mirror.com" before starting OVMS to connect to the HF Hub.
+> **Note:** Kokoro voices are loaded from the `voices/` directory in the model repository and can be selected per request with the `voice` field.
 
-The default configuration should work in most cases but the parameters can be tuned via `export_model.py` script arguments. Run the script with `--help` argument to check available parameters and see the [T2s calculator documentation](../../docs/speech_generation/reference.md) to learn more about configuration options and limitations.
+See the [T2s calculator documentation](../../docs/speech_generation/reference.md) to learn more about configuration options and limitations.
 
 ### Deployment
 
+:::{dropdown} **Deploying with Docker**
+
 **CPU**
 
-Running this command starts the container with CPU only target device:
+Running this command starts the container with CPU target device:
 ```bash
-mkdir -p models
-docker run -d -u $(id -u):$(id -g) --rm -p 8000:8000 -v $(pwd)/models:/models:rw openvino/model_server:latest --rest_port 8000 --model_path /models/microsoft/speecht5_tts --model_name microsoft/speecht5_tts
+mkdir models
+docker run -d -u $(id -u):$(id -g) --rm -p 8000:8000 -v $(pwd)/models:/models:rw openvino/model_server:latest --rest_port 8000 --source_model luis-castillo/Kokoro-82M-OpenVINO-FP16-OVMS --model_repository_path /models --model_name Kokoro-82M-OpenVINO-FP16-OVMS --target_device CPU --task text2speech
 ```
 
-**Deploying on Bare Metal**
+**GPU**
 
-```bat
+In case you want to use GPU device, add extra docker parameters `--device /dev/dri --group-add=$(stat -c "%g" /dev/dri/render* | head -n 1)` and use the GPU image:
+```bash
 mkdir models
-ovms --rest_port 8000 --model_path models/microsoft/speecht5_tts --model_name microsoft/speecht5_tts
+docker run -d -u $(id -u):$(id -g) --rm -p 8000:8000 --device /dev/dri --group-add=$(stat -c "%g" /dev/dri/render* | head -n 1) -v $(pwd)/models:/models:rw openvino/model_server:latest-gpu --rest_port 8000 --source_model luis-castillo/Kokoro-82M-OpenVINO-FP16-OVMS --model_repository_path /models --model_name Kokoro-82M-OpenVINO-FP16-OVMS --target_device GPU --task text2speech
 ```
 
-### Request Generation 
-
-:::{dropdown} **Unary call with curl with default voice**
-
+**NPU**
 
+In case you want to use NPU device, add extra docker parameters `--device /dev/accel --group-add=$(stat -c "%g" /dev/dri/render* | head -n 1)` and use the GPU image:
 ```bash
-curl http://localhost:8000/v3/audio/speech -H "Content-Type: application/json" -d "{\"model\": \"microsoft/speecht5_tts\", \"input\": \"The quick brown fox jumped over the lazy dog\"}" -o speech.wav
+mkdir models
+docker run -d -u $(id -u):$(id -g) --rm -p 8000:8000 --device /dev/accel --group-add=$(stat -c "%g" /dev/dri/render* | head -n 1) -v $(pwd)/models:/models:rw openvino/model_server:latest-gpu --rest_port 8000 --source_model luis-castillo/Kokoro-82M-OpenVINO-FP16-OVMS --model_repository_path /models --model_name Kokoro-82M-OpenVINO-FP16-OVMS --target_device NPU --task text2speech
 ```
 :::
 
-:::{dropdown} **Unary call with OpenAI python library with default voice**
-
-```python
-from pathlib import Path
-from openai import OpenAI
-
-prompt = "The quick brown fox jumped over the lazy dog"
-filename = "speech.wav"
-url="http://localhost:8000/v3"
-
-
-speech_file_path = Path(__file__).parent / "speech.wav"
-client = OpenAI(base_url=url, api_key="not_used")
+:::{dropdown} **Deploying on Bare Metal**
 
-with client.audio.speech.with_streaming_response.create(
-  model="microsoft/speecht5_tts",
-  voice=None,
-  input=prompt
-) as response:
-  response.stream_to_file(speech_file_path)
+**CPU**
+```bat
+mkdir c:\models
+ovms --rest_port 8000 --source_model luis-castillo/Kokoro-82M-OpenVINO-FP16-OVMS --model_repository_path c:\models --model_name Kokoro-82M-OpenVINO-FP16-OVMS --target_device CPU
+```
 
+**GPU**
+```bat
+mkdir c:\models
+ovms --rest_port 8000 --source_model luis-castillo/Kokoro-82M-OpenVINO-FP16-OVMS --model_repository_path c:\models --model_name Kokoro-82M-OpenVINO-FP16-OVMS --target_device GPU
+```
 
-print("Generation finished")
+**NPU**
+```bat
+mkdir c:\models
+ovms --rest_port 8000 --source_model luis-castillo/Kokoro-82M-OpenVINO-FP16-OVMS --model_repository_path c:\models --model_name Kokoro-82M-OpenVINO-FP16-OVMS --target_device NPU
 ```
 :::
 
+### Request Generation 
+
 :::{dropdown} **Unary call with curl**
 
 
 ```bash
-curl http://localhost:8000/v3/audio/speech -H "Content-Type: application/json" -d "{\"model\": \"microsoft/speecht5_tts\", \"voice\":\"voice1\", \"input\": \"The quick brown fox jumped over the lazy dog\"}" -o speech.wav
+curl http://localhost:8000/v3/audio/speech -H "Content-Type: application/json" -d "{\"model\": \"Kokoro-82M-OpenVINO-FP16-OVMS\", \"voice\": \"af_alloy\", \"input\": \"The quick brown fox jumped over the lazy dog\"}" -o speech.wav
 ```
 :::
 
@@ -136,8 +100,8 @@ speech_file_path = Path(__file__).parent / "speech.wav"
 client = OpenAI(base_url=url, api_key="not_used")
 
 with client.audio.speech.with_streaming_response.create(
-  model="microsoft/speecht5_tts",
-  voice="voice1",
+  model="Kokoro-82M-OpenVINO-FP16-OVMS",
+  voice="af_alloy",
   input=prompt
 ) as response:
   response.stream_to_file(speech_file_path)
@@ -152,20 +116,25 @@ Play speech.wav file to check generated speech.
 ## Benchmarking speech generation
 An asynchronous benchmarking client can be used to access the model server performance with various load conditions. Below are execution examples captured on Intel(R) Core(TM) Ultra 7 258V.
 
+> **Note:** `RTFx` (Real-Time Factor, inverted) is calculated as `generated_audio_duration / generation_time`.
+> Values greater than `1.0x` mean faster-than-real-time generation, while values below `1.0x` mean slower-than-real-time.
+
 ```console
-pip install -r https://raw.githubusercontent.com/openvinotoolkit/model_server/refs/heads/main/demos/benchmark/v3/requirements.txt
-curl https://raw.githubusercontent.com/openvinotoolkit/model_server/refs/heads/main/demos/benchmark/v3/benchmark.py -o benchmark.py
-python benchmark.py --api_url http://localhost:8000/v3/audio/speech --model microsoft/speecht5_tts --batch_size 1 --limit 100 --request_rate inf --backend text2speech --dataset edinburghcstr/ami --hf-subset ihm --tokenizer openai/whisper-large-v3-turbo --trust-remote-code True
-Number of documents: 100
-100%|████████████████████████████████████████████████████████████████████████████████| 100/100 [01:58<00:00,  1.19s/it]
-Tokens: 1802
-Success rate: 100.0%. (100/100)
-Throughput - Tokens per second: 15.2
-Mean latency: 63653.98 ms
-Median latency: 66736.83 ms
-Average document length: 18.02 tokens
+pip install -r https://raw.githubusercontent.com/openvinotoolkit/model_server/refs/heads/mkulakow/kokoro_improvements/demos/benchmark/v3/requirements.txt
+curl https://raw.githubusercontent.com/openvinotoolkit/model_server/refs/heads/mkulakow/kokoro_improvements/demos/benchmark/v3/benchmark.py -o benchmark.py
+python benchmark.py --api_url http://localhost:8000/v3/audio/speech --model Kokoro-82M-OpenVINO-FP16-OVMS --batch_size 1 --limit 100 --request_rate inf --backend text2speech --dataset edinburghcstr/ami --hf-subset ihm --voice af_alloy
+Number of documents: 1000
+100%|█████████████████████████████████████████████████████████████████████████████████| 1000/1000 [15:53<00:00,  1.05it/s]
+Success rate: 100.0%. (1000/1000)
+Mean latency: 742.48 ms
+Median latency: 662.10 ms
+Mean RTFx: 5.091x
+Median RTFx: 5.173x
 ```
 
+> **Note:** `RTFx` (Real-Time Factor, inverted) is calculated as `generated_audio_duration / generation_time`.
+> Values greater than `1.0x` mean faster-than-real-time generation, while values below `1.0x` mean slower-than-real-time.
+
 ## Transcription
 ### Model preparation
 Many variances of Whisper models can be deployed in a single command by using pre-configured models from [OpenVINO HuggingFace organization](https://huggingface.co/collections/OpenVINO/speech-to-text) and used both for translations and transcriptions endpoints.
@@ -235,7 +204,7 @@ ovms --rest_port 8000 --model_path /models/openai/whisper-large-v3-turbo --model
 The default configuration should work in most cases but the parameters can be tuned via `export_model.py` script arguments. Run the script with `--help` argument to check available parameters and see the [s2t calculator documentation](../../docs/speech_recognition/reference.md) to learn more about configuration options and limitations.
 
 ### Request Generation 
-Transcript file that was previously generated with audio/speech endpoint.
+Transcribe the speech.wav file generated in the [Speech generation](#speech-generation) section.
 
 > **Note:** Streaming responses are supported for `audio/transcriptions`. `audio/translations` does not support streaming.
 
@@ -360,33 +329,13 @@ Average document length: 10.948 tokens
 ```
 
 ## Translation
-To test translations endpoint we first need to prepare audio file with speech in language other than English, e.g. Spanish. To generate such sample we will use finetuned version of microsoft/speecht5_tts model.
-
-**Deploying with Docker**
+To test the translations endpoint we first need to prepare an audio file with speech in a language other than English, e.g. Spanish. To generate such a sample, follow the [Speech generation](#speech-generation) section to deploy Kokoro and then run:
 
-```bash
-mkdir -p models
-
-python export_model.py text2speech --source_model Sandiago21/speecht5_finetuned_facebook_voxpopuli_spanish --weight-format fp16 --model_name speecht5_tts_spanish --config_file_path models/config.json --model_repository_path models --overwrite_models --vocoder microsoft/speecht5_hifigan
-
-docker run -d -u $(id -u):$(id -g) --rm -p 8000:8000 -v $(pwd)/models:/models:rw openvino/model_server:latest --rest_port 8000 --model_path /models/speecht5_tts_spanish --model_name speecht5_tts_spanish
-
-curl http://localhost:8000/v3/audio/speech -H "Content-Type: application/json" -d "{\"model\": \"speecht5_tts_spanish\", \"input\": \"Madrid es la capital de España\"}" -o speech_spanish.wav
-```
-
-**Deploying on Bare Metal**
-
-```bat
-mkdir models
-
-python export_model.py text2speech --source_model Sandiago21/speecht5_finetuned_facebook_voxpopuli_spanish --weight-format fp16 --model_name speecht5_tts_spanish --config_file_path models/config.json --model_repository_path models --overwrite_models --vocoder microsoft/speecht5_hifigan
-
-ovms --rest_port 8000 --model_path models/speecht5_tts_spanish --model_name speecht5_tts_spanish
-
-curl http://localhost:8000/v3/audio/speech -H "Content-Type: application/json" -d "{\"model\": \"speecht5_tts_spanish\", \"input\": \"Madrid es la capital de España\"}" -o speech_spanish.wav
+```console
+curl http://localhost:8000/v3/audio/speech -H "Content-Type: application/json" -d "{\"model\": \"Kokoro-82M-OpenVINO-FP16-OVMS\", \"voice\": \"em_alex\", \"input\": \"Madrid es la capital de España\"}" -o speech_spanish.wav
 ```
 
-### Model preparation
+### Deployment
 Whisper models can be deployed in a single command by using pre-configured models from [OpenVINO HuggingFace organization](https://huggingface.co/collections/OpenVINO/speech-to-text) and used both for translations and transcriptions endpoints.
 Here is an example of OpenVINO/whisper-large-v3-fp16-ov deployment:
 
@@ -427,7 +376,7 @@ ovms --rest_port 8000 --source_model OpenVINO/whisper-large-v3-fp16-ov --model_r
 :::
 
 ### Request Generation 
-Transcript and translate file that was previously generated with audio/speech endpoint.
+Translate the speech_spanish.wav file generated above.
 
 :::{dropdown} **Unary call with cURL**