Verbara.Sdk.VoiceAi.Tts 2.2.1

Verbara.Sdk.VoiceAi.Tts

Text-to-speech providers for Verbara.Sdk.VoiceAi turn-based pipelines. 6 providers, each implementing ISpeechSynthesizer from Verbara.Sdk.VoiceAi. Native AOT, zero reflection, hand-rolled HTTP/WebSocket clients (no vendor SDK dependencies). MIT licensed.

Providers

TTFA = Time-To-First-Audio. Streaming providers begin returning PCM bytes mid-synthesis; batch providers return the full clip in one response. All providers report metrics via the Verbara.Sdk.VoiceAi.Tts Meter (latency histogram, TTFA histogram, request counters, byte throughput tagged by provider name). Health checks (TtsHealthCheck) auto-registered when the synthesizer is added through DI.

Observability — metric catalog

All metrics are emitted on Meter name Verbara.Sdk.VoiceAi.Tts.

The tts.synthesis.ttfa_ms histogram is the key metric for evaluating provider responsiveness in interactive voice agents. Compare across providers using the voiceai.provider tag.

Install

dotnet add package Verbara.Sdk.VoiceAi.Tts

You almost always want Verbara.Sdk.VoiceAi (the orchestration package) too:

dotnet add package Verbara.Sdk.VoiceAi

Quick start (ElevenLabs)

using Verbara.Sdk.VoiceAi.Tts.DependencyInjection;

services.AddElevenLabsSpeechSynthesizer(o =>
{
    o.ApiKey = configuration["ElevenLabs:ApiKey"]!;
    o.VoiceId = "EXAVITQu4vr4xnSDxMaL";   // "Bella" — pick any from your ElevenLabs library
    o.Model = "eleven_flash_v2_5";        // Flash 2.5 for lowest TTFA
});

The synthesizer is now resolvable as ISpeechSynthesizer and registered with the VoiceAi pipeline.

Per-provider DI extensions

Each provider has its own Add*SpeechSynthesizer extension (in Verbara.Sdk.VoiceAi.Tts.DependencyInjection):

services.AddElevenLabsSpeechSynthesizer(o => { ... });
services.AddCartesiaSpeechSynthesizer(o => { ... });
services.AddSpeechmaticsSpeechSynthesizer(o => { ... });
services.AddAzureTtsSpeechSynthesizer(o => { ... });

Choosing a provider

• Best TTFA → Cartesia Sonic-3 (40-90 ms). Lowest perceived latency for interactive AI agents.
• Best quality / familiarity → ElevenLabs (Flash 2.5 is fast; Multilingual v2 is premium).
• Best language coverage → Azure TTS (broad locale support; batch-only).
• Mid-market enterprise → Speechmatics (good balance).

Examples

• Examples/VoiceAiExample/ — ElevenLabs + Deepgram + echo handler (default demo).
• Examples/VoiceAiCartesiaExample/ — Cartesia Sonic-3 with sub-100 ms TTFA.
• Examples/VoiceAiSpeechmaticsExample/ — Speechmatics TTS.

Native AOT

All HTTP/WebSocket clients hand-rolled with HttpClient / ClientWebSocket. JSON serialization via source-generated JsonSerializerContext (VoiceAiTtsJsonContext). 0 trim warnings. See ADR-0014 for the no-vendor-SDK rationale.

License

MIT. Part of the Verbara.Sdk project.