The duration of apodization in seconds.

Repeated field holding individual synthesis requests that make up the batch request.

Repeated field holding individual synthesis responses that correspond to the input requests in the batch.

Flag indicating whether caching is active.

The maximum size of the memory cache.

The sampling rate for caching.

Flag indicating whether to load cache.

Flag indicating whether to save cache.

The directory path to save the cache.

Text-to-mel inference settings.

Mel-to-audio inference settings.

The prefix for the custom phonemizer ID.

Repeated field of Map messages representing word-to-phoneme mappings.

The ID of the custom phonemizer.

Repeated field of Map messages representing word-to-phoneme mappings.

The batch size for inference.

Flag indicating whether to use GPU for inference.

The length scale for inference.

The noise scale for inference.

The path to the GlowTTS model.

Repeated field containing the cleaners for text normalization.

The path to the parameter configuration.

The batch size for inference.

The length scale for inference.

The noise scale for inference.

Repeated field containing the cleaners for text normalization.

The maximum text length allowed.

The path to the parameter configuration.

The name of the Triton model.

The host of the Triton inference server which serves the model.

The port of the Triton inference server which serves the model.

Flag indicating whether to use GPU for inference.

The batch size for inference.

The path to the HiFiGan configuration.

The path to the HiFiGan model.

The path to the HiFiGan Triton configuration.

The name of the Triton model.

The host of the Triton inference server which serves the model.

The port of the Triton inference server which serves the model.

Repeated field of pipeline IDs to filter the list of custom phonemizers.

Repeated field of CustomPhonemizerProto messages representing the custom phonemizers.

Optional. Define the speaker sex.

Optional. Define the pipeline owner/ owners.

Optional. Define the speaker name/ names.

Optional. Define the language/ languages.

Required. Define the domain/ domains that satisfy/ies the specifications in the ListT2sDomainsRequest.

Optional. Define the speaker sex.

Optional. Define the pipeline owner/ owners.

Optional. Define the speaker name/ names.

Optional. Define the domain/ domains.

Required. Define the language/ languages that satisfy/ies the specifications in the ListT2sLanguagesRequest.

Optional. Define the language.

Required. Representation of a list of normalization pipelines configurations. Retrieved by ListT2sNormalizationPipelines, containing the configurations of normalization pipelines with the specifications received in the ListT2sNormalizationPipelinesRequest.

Optional. Define the language/ languages.

Optional. Define the speaker sex.

Optional. Define the pipeline owner/ owners.

Optional. Define the speaker name/ names.

Optional. Define the domain/ domains.

Required. Representation of a list of pipelines configurations. Retrieved by ListT2sPipelines, containing the configurations of pipelines with the specifications received in the ListT2sPipelinesRequest.

The initial noise value.

The window size.

The noise threshold.

The word to be mapped.

The phoneme groups associated with the word.

The path to the MbMelgan Triton configuration.

The path to the MbMelgan statistics.

The name of the Triton model.

The host of the Triton inference server which serves the model.

The port of the Triton inference server which serves the model.

The type of mel-to-audio inference (vocoder), selecting which of the settings below is applied (mb_melgan_triton, hifi_gan or hifi_gan_triton).

MbMelgan Triton inference settings.

HiFiGan inference settings.

HiFiGan Triton inference settings.

The ID of the text-to-speech pipeline.

The text to be normalized.

The normalized text.

The ID of the phonemizer.

The duration of silence in seconds.

Repeated field containing pipeline names.

Logmnse postprocessing settings.

Wiener postprocessing settings.

Apodization postprocessing settings.

Model name for the inference server.

Language for inference server

Path of the embedding file

Host name for the inference server

Port of the inference server

Authorization headers for the inference server

Voice id indicating the speaker name

Model name for the inference server.

Language for inference server

Host name for the inference server

Port of the inference server

Authorization headers for the inference server

Required. Represents the pipeline id of the model configuration that will be used.

Optional. This parameter is used for time stretching which is the process of changing the speed or duration of an audio. It should be much more than 1.0. 0 is not a valid number for this variable. The default value is 1.

Optional. Defines the noise in the generated audio. It should be between 0.0 and 1. The default value is 0.0

Optional. Defines the sample rate of the generated wav file. The default value is 22050.

Optional. Defines the pulse-code modulation of the wav file. The default value is PCM_16.

Optional. Defines the format of the desired audio. The default value is wav.

Optional. Define if cache should be used or not. The default value is False.

Optional. t2s_service_config provides the configuration of the service such as API key, bearer tokens, JWT, and other header information as key value pairs, e.g.,

MY_API_KEY='LKJDIFe244LKJOI'

• A1. aws_access_key_id (required) Access key id to access Amazon WEB Service.
• A2. aws_secret_access_key (required) Secret access key to access Amazon WEB Service.
• A3. region (required) Region name of Amazon Server.

t2s_config_service={'aws_access_key_id': 'YOUR_AWS_ACCESS_KEY_ID', 'aws_secret_access_key': 'YOUR_AWS_SECRET_ACCESS_KEY', 'region': 'YOUR_AMAZON_SERVER_REGION_NAME'}

• B1. api_key (required) API key of ElevenLabs cloud provider to access its T2S service.

t2s_config_service={'api_key': 'YOUR_ELEVENLABS_API_KEY'}

• C1. api_key (required) API key of Google cloud provider to access its T2S service.
• C2. api_endpoint (optional) Regional API endpoint of Google cloud T2S service. (Defaults to 'eu-texttospeech.googleapis.com')

t2s_config_service={'api_key': 'YOUR_GOOGLE_CLOUD_API_KEY', 'api_endpoint': 'YOUR_GOOGLE_CLOUD_API_ENDPOINT'}

• D1. subscription_key (required) Subscription key to access Microsoft Azure Service.
• D2. region (required) Region name of Microsoft Azure Server.

t2s_config_service={'subscription_key': 'YOUR_MICROSOFT_AZURE_SUBSCRIPTION_KEY', 'region': 'YOUR_MICROSOFT_AZURE_SERVER_REGION_NAME'}

Optional. Defines the cloud provider's specific configuration for using text to speech cloud services The default value is None.

Optional. Define t2s_normalization config parameters for this specific request. The default values are set in the config file and the values set via RequestConfig are set just for this specific request and will not update the pipeline.

Optional. Define a dict which specifies the phonemes for a special word.

Optional. Define an instruction or prompt to be passed to the TTS/LLM backend for this request.

This field is only honored by the qwen3-tts-voicedesign model, where it is a natural-language description of the target voice to design, e.g.

a calm female voice with a slight British accent

It is ignored by every other pipeline, including the qwen3-tts CustomVoice and base models and the local vits / glow-tts models. To control the delivery / emotion of the qwen3-tts CustomVoice model, embed the control tags in the request text instead (see SynthesizeRequest.text), not via this field.

Text-to-audio inference settings.

Required. Represents the text that will be transformed to speech. All the properties according to the input text in SynthesizeRequest can be also applied here.

Required. Represents the specifications needed to do the text to speech transformation.

Required. The unique identifier (UUID) assigned to the generated audio.

Required. Generated file with the parameters described in request.

Required. Time taken to generate the audio, in seconds.

Required. Length (duration) of the generated audio, in seconds.

Required. Text from which audio was generated.

Required. Configuration from which audio was generated.

Optional. Normalized text.

Optional. Sampling rate of the generated audio, in Hertz (Hz).

Required. Represents the text that will be transformed to speech.

The text may be plain text, or it may embed SSML markup, phonemes and voice-control tags to shape how it is spoken. The sections below list every supported form with a copy-paste example. Which forms take effect depends on the pipeline (model) selected via RequestConfig.t2s_pipeline_id; unsupported tags are stripped and the plain text is still spoken.

1. Plain text

• Simple text:

  Hello, how are you?

2. SSML <say-as> tags (local vits / qwen3-tts and cloud providers)

Rewrites the enclosed text into its spoken form (spelled out letters, digits, symbols, ...). Supported by the local normalization pipeline and forwarded to the cloud providers.

• Phone:

  <say-as interpret-as="phone">+12354321</say-as>

• Email:

  <say-as interpret-as="email">voices@ondewo.com</say-as>

• URL:

  <say-as interpret-as="url">ondewo.com/en/</say-as>

• Spell (spell out each character):

  <say-as interpret-as="spell">AP732</say-as>

• Spell with names (e.g. "A like Anton"):

  <say-as interpret-as="spell-with-names">AHO32</say-as>

• Callsigns short:

  <say-as interpret-as="callsign-short">AUA439</say-as>

• Callsigns long:

  <say-as interpret-as="callsign-long">AAL439</say-as>

3. Pauses / breaks

Inserts a pause of the given duration into the spoken text. Supported by the local pipelines and forwarded to the cloud providers. The duration is written either with a unit - ms for milliseconds, s for seconds - or as a bare number, which is read as seconds: <break time="0.5"/> is half a second. A bare number is therefore never milliseconds; <break time="2000"/> asks for 2000 seconds and is rejected with an error - write <break time="2000ms"/> instead. A single break may not be longer than 60 seconds.

• Milliseconds:

  Welcome to our service. <break time="500ms"/> How can I help you today?

• Seconds:

  Please hold the line. <break time="2s"/> Thank you for waiting.

• Seconds without a unit:

  Let me check that for you. <break time="1.5"/> I found your booking.

4. Phonemes / pronunciation control

The phoneme alphabet depends exclusively on the model type:

4a. IPA phonemes (vits and cloud providers only) - the vits models and the cloud providers (Amazon Polly, Microsoft Azure and Google Cloud TTS) control pronunciation with the SSML <phoneme> element using the IPA alphabet (alphabet="ipa"). The IPA string goes in the ph attribute and the written word between the tags:

• English:

  I would like a <phoneme alphabet="ipa" ph="təˈmeɪtoʊ">tomato</phoneme> please

• German:

  Ich hätte gerne ein <phoneme alphabet="ipa" ph="ˈbaɪʃpiːl">Beispiel</phoneme> bitte

4b. Arpabet phonemes (glow model only) - the glow model controls pronunciation with Arpabet phonemes, supplied through the same SSML <phoneme> element but with alphabet="arpabet". The ARPAbet transcription goes in the ph attribute and the written word between the tags (the phonemes are not written in curly braces {} in the request text):

• English:

  Please say <phoneme alphabet="arpabet" ph="T EH1 S T">test</phoneme> out loud

• German:

  Ich sage <phoneme alphabet="arpabet" ph="T EH1 S T">test</phoneme> laut

4c. qwen3-tts has no phonemes - use the verbalized form instead - the qwen3-tts models do not support Arpabet or IPA phonemes; the <phoneme> element and its ph string are ignored. Instead, to have a special text (symbols, codes, abbreviations, ...) spoken correctly, pass its verbalization: the plain, written-out spoken form of the text in the target language. For example, to have A-7+8* pronounced correctly:

• English:

  ai dash seven plus eight asterisk

• German:

  ah bindestrich sieben plus acht sternchen

5. Voice control for qwen3-tts (style / prosody)

These ONDEWO-namespaced tags are compiled into the model instruction and stripped from the spoken text, so the transcription stays the plain sentence while the delivery changes. They only take effect on the qwen3-tts CustomVoice model; on the base and voicedesign models they are stripped from the text without changing the delivery.

qwen3-tts is steered by a natural-language instruction rather than by numeric parameters, so every dimension has exactly one step per direction and the values listed below are the whole vocabulary. Any other value is reported in the server log and then ignored.

• Style - one of the four primary styles apologetic, calm, empathetic, firm, or one of the accepted synonyms cheerful / happy / joyful, sad / sorrowful, angry / furious, excited. An unknown name falls back to neutral, which adds no style at all, and a warning is logged:

  <ondewo:style name="calm">please stay calm everything will be fine</ondewo:style>

• Prosody - rate (slow / medium / fast) and volume (soft / medium / loud), where medium means "leave it alone" and adds nothing.

  <prosody rate="slow" volume="loud">hello there</prosody>

• Nested combination (style + prosody):

  <ondewo:style name="calm"><prosody rate="slow" volume="soft">hello there</prosody></ondewo:style>

• Free-form instruction (CustomVoice only) - describe the delivery in natural language in the instruction attribute. The attribute value is used verbatim as the model instruction and takes precedence over the style / prosody tags; the wrapped inner text is kept as the spoken text and the instruction should always be written in English:

  <ondewo:instruct instruction="speak slowly in a warm and reassuring tone">hello there</ondewo:instruct>

• Inline style tags (CustomVoice only) - the same style vocabulary written positionally in square brackets, so [calm] hello there and the <ondewo:style name="calm"> element compile to the identical instruction. Such a tag applies to everything that follows it until another tag or the [Neutral] reset replaces it.

  [excited] we shipped it. it works. [calm] now let us review.

Styles supported per voice

A style is rendered by the voice, not by the markup, and no voice renders all of them. The table below lists what each qwen3-tts speaker supports; the speaker is fixed by the pipeline selected through RequestConfig.t2s_pipeline_id.

The table applies to both ways of asking for a style. The <ondewo:style name="calm"> element and the inline [calm] tag compile to the very same instruction, so they are limited by the voice in exactly the same way - switching from one spelling to the other never makes a style work that the voice cannot render.

The server does not check your request against this table. If you ask a voice for a style it cannot render, the request still succeeds: audio is returned, no error is raised and nothing is written to the log. The only symptom is that the voice sounds exactly as it would have without the style. Use the table to choose a suitable voice before sending the request.

This is not the same as writing a style name that does not exist, which is reported: <ondewo:style name="bored"> logs a warning and falls back to no style, because bored is outside the vocabulary listed further above. The silent case is a name that is in the vocabulary but beyond what the selected voice can do, for example excited on dylan.

Notes on the table:

• dylan and serena render no style at all - use them for neutral delivery only, and see the prosody table below for what little can still be varied on them.
• ryan and vivian cover the widest range, so they are the safest choice when the style is chosen at runtime rather than fixed in advance.
• A synonym behaves exactly like its primary name, so happy / joyful work wherever cheerful does, as do sorrowful for sad and furious for angry.

Prosody supported per voice

Rate and volume are limited per voice in the same way, and the same caveat applies: the server does not check this table either, so asking for a step a voice cannot render succeeds silently and leaves the delivery unchanged. Only the two steps of each dimension are listed - medium (equally default) is available on every voice, because it means "leave this dimension alone" and sends no instruction at all.

Notes on the table:

• uncle_fu does not respond to rate at all - change its pace by rewriting the text, or pick another voice.
• dylan renders no style but does respond to rate and to volume="soft", so <prosody> is the only delivery control that works on it.
• serena is the most constrained voice: rate="slow" is the one control it responds to, and no style, no rate="fast" and no volume step has any effect. Treat it as a fixed delivery that can only be slowed down.
• ono_anna, ryan, sohee and aiden render every step of both dimensions.

Every control is positional - it colours exactly the text it wraps - and a delivery change may only land on a sentence boundary. Markup wrapping part of a sentence is widened to the whole sentence, because a sentence fragment sent as a complete utterance makes the model hallucinate. Each delivery change also costs one additional inference call, so marking up whole sentences, or leaving the text unmarked, synthesises fastest.

See also RequestConfig.instruction for supplying the instruction / voice description directly instead of embedding it in the text.

6. Provider-specific SSML (cloud pipelines only)

Each cloud provider documents its own SSML vocabulary; use the elements matching the provider behind the selected pipeline. Standard <prosody>, <emphasis> and <break> are common to all three.

• Google Cloud TTS:

  <google:style name="calm">please stay calm everything will be fine</google:style>

  More information: Google Cloud TTS - SSML styles
• Microsoft Azure:

  <mstts:express-as style="cheerful">please stay calm everything will be fine</mstts:express-as>

  More information: Microsoft Azure - speech synthesis markup, voice and speaking styles
• Amazon Polly (whispered):

  <amazon:effect name="whispered">please stay calm everything will be fine</amazon:effect>

  More information: Amazon Polly - whispered speech with amazon:effect
• Amazon Polly (soft phonation):

  <amazon:effect phonation="soft">please stay calm everything will be fine</amazon:effect>

Required. Represents the specifications needed to do the text to speech transformation.

Required. The unique identifier (UUID) assigned to the generated audio.

Required. Generated file with the parameters described in request.

Required. Time taken to generate the audio, in seconds.

Required. Length (duration) of the generated audio, in seconds.

Required. Text from which audio was generated.

Required. Configuration from which audio was generated.

Optional. Normalized text.

Optional. Sampling rate of the generated audio, in Hertz (Hz).

The custom length scale for general text.

The custom length scale for email text.

The custom length scale for URL text.

The custom length scale for phone number text.

The custom length scale for spelled-out text.

The custom length scale for spelled-out text with names.

The custom length scale for long callsigns.

The custom length scale for short callsigns.

The language supported by the service.

The speaker's sex or gender.

The owner of the text-to-speech pipeline.

Additional comments or notes.

The name of the speaker.

The domain or context of the service.

version number

The type of inference, selecting which of the settings below is applied: composite_inference (two-stage text-to-mel then mel-to-audio) or single_inference (single-stage text-to-audio).

Composite (two-stage) inference settings, used when type selects composite inference.

Single (one-stage) inference settings, used when type selects single inference.

Caching settings.

The language for which the normalization is applied.

The pipeline(s) used for normalization.

The ID of the custom phonemizer, if used.

Custom length scales for different text types.

The mapping for Arpabet phonemes.

The mapping for numeric expressions.

The mapping for callsigns.

The mapping for phoneme correction.

Configuration for Eleven Labs text-to-speech provider.

Configuration for Google text-to-speech provider.

Configuration for Microsoft text-to-speech provider.

Stability level for inference, influencing consistency of generated speech. It is in the range [0.0, 1.0].

Boost value for similarity to enhance the similarity of the generated voice to a target voice. It is in the range [0.0, 1.0].

Style parameter to control the expression or emotion in speech. It is in the range [0.0, 1.0].

Enables or disables speaker boost for emphasis on clarity and loudness.

Specifies type of text normalization to apply during processing. Available options are 'auto', 'on', and 'off'.

Speaking rate for inference, controlling the speed of generated speech. It is in the range [0.25, 4.0].

Volume gain in dB applied to the generated speech. It is in the range [-96.0, 16.0].

Pitch adjustment for inference, allowing control over voice pitch. It is in the range in the range [-20.0, 20.0].

Determines whether to use the default speaker voice.

Voice ID indicating the speaker

Model id for the inference server.

Language of the generated audio. It should be 4-Letter language code.

Model ID indicating the name of the model

Voice ID indicating the speaker

Voice setting of the inference

Flag to indicate applying text normalization

Voice ID indicating the speaker

Speaking rate to control the speed of audio.

Volume gain in db to control volume of the audio.

pitch value of the audio

speaker_language indicating the speaker language code

Voice ID indicating the speaker.

Flag to indicate using the default speaker.

Required. Defines the id of the pipeline.

The type of text-to-audio inference, selecting which of the settings below is applied: a local model (vits, vits_triton) or a cloud / neural service (t2s_cloud_service_elevenlabs, t2s_cloud_service_amazon, t2s_cloud_service_google, t2s_cloud_service_microsoft, qwen3_tts_custom_voice, qwen3_tts_base).

Vits inference settings.

Vits Triton inference settings.

ElevenLabs cloud service inference settings.

Amazon cloud service inference settings.

Google cloud service inference settings.

Microsoft cloud service inference settings.

Qwen3-TTS-custom-voice inference settings.

Qwen3-TTS-base inference settings for voice cloning.

The type of text-to-mel inference, selecting which of the settings below is applied (glow_tts or glow_tts_triton).

GlowTTS inference settings.

GlowTTS Triton inference settings.

Required. Defines the id of the pipeline.

Required. Defines the description of the pipeline representation.

Required. Defines if the pipeline is active or inactive.

Required. Defines the inference of the pipeline representation.

Required. Defines the normalization process of the pipeline representation.

Required. Defines the postprocessing process of the pipeline representation.

The ID of the custom phonemizer to be updated.

The update method.

Repeated field of Map messages representing word-to-phoneme mappings.

The batch size for inference.

Flag indicating whether to use GPU for inference.

The length scale for inference.

The noise scale for inference.

The path to the Vits model.

Repeated field containing the cleaners for text normalization.

The path to the parameter configuration.

The batch size for inference.

The length scale for inference.

The noise scale for inference.

Repeated field containing the cleaners for text normalization.

The maximum text length allowed.

The path to the parameter configuration.

The name of the Triton model.

The host of the Triton inference server which serves the model.

The port of the Triton inference server which serves the model.

Required. The sample audio of the speaker whose voice should be cloned.

Required. The transcription of the sample audio.

Required. The name of the speaker to associate with the cloned voice.

Required. The language of the speaker in the sample audio.

Required. The name of the model used for voice cloning.

stability value for elevenlabs inference

similarity boost value for ElevenLabs inference.

style boost value for ElevenLabs inference.

Flag to indicate speaker boost

The frame length.

The LPC order.

The number of iterations.

The alpha value.

The threshold value.