Text-to-Speech (TTS) Providers

dIKta.me implements a flexible, multi-tier execution strategy for Text-to-Speech (TTS) generation, enabling offline local inference via ONNX and high-quality cloud generation through a unified provider interface.

Provider Architecture

All TTS engines implement the ITTSProvider interface, allowing the TtsSpeaker service to route requests seamlessly.

public interface ITTSProvider : IDisposable
{
    string ProviderName { get; }
    bool SupportsStreaming { get; }

    /// <summary>Synthesize text to audio bytes (full generation).</summary>
    Task<TtsResult> SynthesizeAsync(
        string text,
        string? voiceId = null,
        CancellationToken cancellationToken = default);

    /// <summary>Check if the provider is ready (model loaded / API reachable).</summary>
    Task<bool> IsAvailableAsync(CancellationToken cancellationToken = default);
}

The TtsResult encapsulates the synthesized audio along with metadata, including the PCM/WAV byte array, sample rate, and generation latency.

Available Providers

KokoroTtsProvider (Local)

The default local provider utilizes KokoroSharp, wrapping the highly efficient Kokoro-ONNX models (82M parameters).

• Execution: Runs in-process with zero external dependencies.
• Variants: Supports int8 (88 MB), fp16 (169 MB), and fp32 (310 MB) precision models via Microsoft.ML.OnnxRuntime.
• Latency: Generates speech typically in 300–500ms.
• License compliance: Kokoro's internal consumption of the GPLv3 eSpeak-NG phonemizer is isolated to process invocations under "mere aggregation," requiring only attribution.

DeepgramTtsProvider (Cloud Default)

Leverages the existing Deepgram .NET SDK integration to execute the Aura-2 TTS models using the same API key deployed for the primary dictate STT operations. Provides 90-200ms TTFB across HTTP/WebSockets.

InworldTtsProvider & OpenAITtsProvider (Premium Cloud)

Implements standard REST clients for external HTTP integrations using standard ASP.NET HttpClient patterns. Authentication keys are dynamically injected via DPAPI SecureStorage.

Lifecycle and Routing

The TTSProviderFactory initializes and caches ITTSProvider instances to prevent redundant allocations or constant re-loading of local ONNX models.

During playback requests, the TTSRouter evaluates the active application TtsSettings. It dynamically falls back to alternate providers if the chosen API is unavailable or if the local model has failed to initialize.

Audio Player Pipeline

Audio is subsequently streamed directly through the TtsPlayerService leveraging NAudio's WasapiOut. The player instance seamlessly synchronizes with the AudioDucker to drop the system volume of background applications during active playback. Playback can be interrupted instantaneously by evaluating hotkey states or application state transitions (e.g. initiating a new dictation instance).

Text Sanitization

Raw Markdown from LLM responses is heavily sanitized via the static TextCleaner class before processing. This strips header (#) markers, transforms list bullets into pauses, expands common symbols ($ to "dollars"), and aggressively truncates excessively long output boundaries prior to inference.