nvidia

type Endpointing struct {
	StartHistory     *int32
	StartThreshold   *float32
	StopHistory      *int32
	StopThreshold    *float32
	StopHistoryEOU   *int32
	StopThresholdEOU *float32
}

type STTConfig struct {
	// Server is the Riva gRPC endpoint; empty uses NVIDIA's hosted endpoint. For
	// a local NIM this is typically "localhost:50051".
	Server string
	// APIKey is the NVIDIA API key, sent as an "authorization: Bearer" header.
	// Required by the hosted endpoint; omit for an unauthenticated local NIM.
	APIKey string
	// FunctionID is the NVIDIA Cloud Function id, sent as a "function-id" header.
	// Required by the hosted endpoint to select the model; omit for a local NIM.
	// It identifies one deployment of a model rather than the model itself, so
	// it changes whenever NVIDIA redeploys; the current one is on the model's
	// page under build.nvidia.com.
	FunctionID string
	// Model selects a served model by name; empty lets the server choose. A local
	// Riva build usually serves one model, so this can stay empty.
	Model string
	// Language of the audio, mapped to a Riva BCP-47 code; the zero value uses US
	// English.
	Language language.Language
	// SampleRate is the input audio sample rate; 0 uses the transport's rate.
	SampleRate int

	// UseSSL wraps the gRPC connection in TLS; nil defaults to true. Set to a
	// pointer to false for a plaintext local NIM.
	UseSSL *bool
	// InterimResults requests tentative partial transcripts; nil defaults to true.
	InterimResults *bool
	// AutomaticPunctuation adds punctuation to results; nil defaults to true.
	AutomaticPunctuation *bool
	// VerbatimTranscripts disables inverse text normalization; nil defaults to
	// true.
	VerbatimTranscripts *bool
	// ProfanityFilter masks profanities in the transcript.
	ProfanityFilter bool
	// MaxAlternatives caps the hypotheses returned per result; 0 uses 1.
	MaxAlternatives int
	// AudioChannelCount is the input channel count; 0 uses 1 (mono).
	AudioChannelCount int
	// CustomConfiguration passes request-level options to the model pipeline
	// (for example "enable_vad_endpointing:true"), keyed by option name.
	CustomConfiguration map[string]string
	// Endpointing tunes the server-side utterance detector; nil uses the server
	// defaults.
	Endpointing *Endpointing

	// TTFSP99 overrides the measured transcript latency the turn strategies
	// size their wait by; 0 uses stt.NvidiaTTFSP99.
	TTFSP99 time.Duration
}

type SegmentedSTTConfig struct {
	// Server is the Riva gRPC endpoint; empty uses NVIDIA's hosted endpoint. For
	// a local NIM this is typically "localhost:50051".
	Server string
	// APIKey is the NVIDIA API key, sent as an "authorization: Bearer" header.
	// Required by the hosted endpoint; omit for an unauthenticated local NIM.
	APIKey string
	// FunctionID is the NVIDIA Cloud Function id, sent as a "function-id" header.
	// Required by the hosted endpoint to select the model; omit for a local NIM.
	// It identifies one deployment of a model rather than the model itself, so
	// it changes whenever NVIDIA redeploys; the current one is on the model's
	// page under build.nvidia.com.
	FunctionID string
	// Model selects a served model by name; empty lets the server choose.
	Model string
	// Language of the audio, mapped to a Riva BCP-47 code; the zero value uses US
	// English.
	Language language.Language
	// SampleRate is the input audio sample rate; 0 uses the transport's rate.
	SampleRate int
	// UseSSL wraps the gRPC connection in TLS; nil defaults to true. Set to a
	// pointer to false for a plaintext local NIM.
	UseSSL *bool
	// AutomaticPunctuation adds punctuation to results; nil defaults to true.
	AutomaticPunctuation *bool
	// VerbatimTranscripts disables inverse text normalization; nil defaults to
	// true.
	VerbatimTranscripts *bool
	// ProfanityFilter masks profanities in the transcript.
	ProfanityFilter bool
	// MaxAlternatives caps the hypotheses returned per result; 0 uses 1.
	MaxAlternatives int
	// AudioChannelCount is the input channel count; 0 uses 1 (mono).
	AudioChannelCount int
	// CustomConfiguration passes request-level options to the model pipeline,
	// keyed by option name.
	CustomConfiguration map[string]string

	// TTFSP99 overrides the measured transcript latency the turn strategies
	// size their wait by; 0 uses stt.NvidiaTTFSP99.
	TTFSP99 time.Duration
}

type TTSConfig struct {
	// Server is the Riva gRPC endpoint; empty uses NVIDIA's hosted endpoint. For
	// a local NIM this is typically "localhost:50051".
	Server string
	// APIKey is the NVIDIA API key, sent as an "authorization: Bearer" header.
	// Required by the hosted endpoint; omit for an unauthenticated local NIM.
	APIKey string
	// FunctionID is the NVIDIA Cloud Function id, sent as a "function-id" header.
	// It selects the model on the hosted endpoint; empty uses the multilingual
	// Magpie function. Set it to "-" to send no function id, as a local NIM
	// expects.
	FunctionID string
	// Model names the served model. It labels the metrics and traces and is what
	// a cost-tracking backend prices against; empty uses the hosted default. It
	// does not select the model, FunctionID does.
	Model string
	// Voice is the voice name; empty uses a multilingual Magpie voice.
	Voice string
	// Language for synthesis, mapped to a Riva BCP-47 code; the zero value uses
	// US English.
	Language language.Language
	// SampleRate is the PCM rate requested from the server and emitted
	// downstream; 0 uses 24 kHz. Rates below 8 kHz do not produce usable audio.
	SampleRate int
	// UseSSL wraps the gRPC connection in TLS; nil defaults to true. Set to a
	// pointer to false for a plaintext local NIM.
	UseSSL *bool
	// CustomDictionary maps a written form to its IPA pronunciation, for example
	// {"NVIDIA": "ɛn.vɪ.diː.ʌ"}; empty sends none.
	CustomDictionary map[string]string
	// CustomConfiguration passes model-specific options the schema does not name
	// (for example "exaggeration_factor"), keyed by option name.
	CustomConfiguration map[string]string
	// ZeroShot supplies the voice-cloning prompt a zero-shot model needs; nil
	// omits it, which is what every other model expects.
	ZeroShot *ZeroShot `validate:"omitempty"`
}

type ZeroShot struct {
	// AudioPrompt is the prompt audio. NVIDIA recommends 16-bit mono, 22.05 kHz
	// or higher, and 3 to 10 seconds long. Required.
	AudioPrompt []byte `validate:"required,min=1"`
	// Encoding of the prompt, either "pcm" (the default) or "oggopus".
	Encoding string `validate:"omitempty,oneof=pcm oggopus"`
	// SampleRate of the prompt; 0 lets the server assume 22050.
	SampleRate int
	// Quality is how many times the prompt passes through the decoder, 1 to 40;
	// 0 uses 20.
	Quality int `validate:"omitempty,min=1,max=40"`
	// Transcript of the prompt audio; empty sends none.
	Transcript string
}