Behind Nock9 min read
Why Nock Uses xAI and Speechify for Voice Agents
Building a good voice agent comes down to a few things people notice immediately: how quickly it responds, whether it is pleasant to listen to, whether it can do something useful, and what it costs to keep running.
Read the story
On this page 9 sections
For Nock, xAI and Speechify are two of our best choices for that balance. Both are part of what Nock is built with. The reason is fairly simple: they offer useful voice capabilities, good audio options and pricing that makes sense to build around.
I think Speechify's audio quality is especially good. That is my preference, not a claim that one provider wins every listening test. But when you are building something people will hear throughout the day, liking the voice is not a small detail.
If you are choosing a voice agent API or a text-to-speech API, these are two providers I would put on the shortlist. Here is why.
The short answer: good voices at a workable cost
A voice agent does not need the most expensive voice available. It needs a voice that people understand, a response time that does not interrupt the flow of conversation, and economics that still work when people actually use it.
That is what I mean by relatively affordable, or "cheapish." Not free. Not necessarily the lowest price in every comparison. Just a sensible balance between what the system costs and what the person using it gets back.
Speechify is particularly interesting when the spoken output matters and you want a dedicated text-to-speech service. xAI is interesting because its voice offering goes beyond generating audio: it also offers real-time speech-to-speech conversation and tool calling.
Those are related capabilities, but they are not the same product.
| What you need | Why Speechify is worth considering | Why xAI is worth considering |
|---|---|---|
| A voice for an existing AI assistant | A dedicated TTS layer that can speak the text your assistant produces. | A standalone TTS API as well as a broader voice stack. |
| A real-time conversational agent | Fits into an agent platform with your chosen speech recognition and language model. | A native speech-to-speech API with turn detection and tools. |
| Audio that suits your product | A catalog of voices and streaming synthesis; Speechify's sound is my personal preference here. | A range of expressive voices shared across its TTS and real-time APIs. |
| Costs you can plan around | Developer plans with included usage and per-character rates. | Published per-character TTS pricing and per-minute speech-to-speech pricing. |
These distinctions come from the providers' Speechify API documentation and xAI voice documentation. The preference for Speechify's sound is mine.
Why Speechify is a strong text-to-speech choice
The simplest way to think about Speechify's developer API is that it supplies the speaking voice. Your application produces text, and Speechify turns that text into audio.
That is useful if you already have an assistant, an LLM or an agent framework you want to keep. You do not have to choose the reasoning model and the speaking voice as one inseparable package.
Speechify documents streaming synthesis, pronunciation controls, voice cloning with verified consent, and integrations with frameworks such as LiveKit and Pipecat. Its current recommended model, Simba 3.5, documents support for 48 languages. Language support and the right accent are different questions, so it still makes sense to audition a voice for the audience you actually have. Speechify Build documentation
What interests me most is the sound. A voice can be technically clear and still not be something you want to hear all day. Pacing, pronunciation and how it handles an ordinary sentence matter just as much as whether a short demo sounds impressive.
For Nock, that everyday experience matters. Nock is a voice agent you use while working, not just an audio clip you play once. It can take dictation, answer questions and hand larger jobs to background workers. The speaking voice belongs to that experience. What Nock is
That is why I think Speechify deserves serious consideration when choosing a text-to-speech API for a voice assistant.
Why xAI is a strong voice agent choice
xAI is appealing for a slightly different reason: it offers a broader set of voice capabilities in one ecosystem.
Its public APIs include text to speech, speech to text and real-time speech-to-speech conversation. The real-time API supports bidirectional audio, voice activity detection and tools. That makes it relevant to an agent that needs to respond and take useful actions, rather than only read prepared text aloud. xAI speech-to-speech documentation
For a developer, that gives you options. You can evaluate xAI as a native conversational system, or evaluate its individual speech services as components in a larger application.
It does not remove the work of building a product. You still need permissions, reliable tool integrations, sensible failure handling and a clear interface. A voice agent that sounds good but performs the wrong action is not a good voice agent.
But xAI gives developers a practical starting point for spoken conversation with tool use. Combined with published usage pricing, that makes it one of the providers I would consider early rather than late.
What "affordable" actually means for a voice agent
The API price is important, but it is not the whole bill. Speech generation, transcription, the language model, tools, hosting and any telephone infrastructure can all contribute to the cost.
Here is a dated snapshot of the published voice pricing, checked on 15 November 2026. These are provider rates, not Nock's retail pricing or a claim about our actual operating costs.
| Service | Published pricing snapshot | What to watch |
|---|---|---|
| xAI standalone TTS | $15 per million input characters. | A TTS charge is not the cost of a complete agent. |
| xAI Grok Voice 3.0 speech to speech | $0.08 per minute, listed as $4.80 per hour, plus $0.004 per text input. | Tool calls and other services may add charges; check how your intended usage is billed. |
| Speechify Simba 3.5 TTS | Starter, Pro and Scale list $15, $12 and $10 per million characters respectively. | The checked page temporarily bills those rates at $10, $8 and $6 until 21 November 2026. Subscription fees, included balances and top-ups affect effective cost. |
Check the current xAI pricing and Speechify developer pricing before building a budget. Speechify's consumer Reader subscription is separate from API usage.
The important question is not just "Which API is cheapest?" It is "Can I deliver an experience people want to use at a cost the product can support?"
For Nock, the balance of cost and voice quality is why these two providers make sense to us. That is a product decision, not a universal price ranking.
Speechify versus xAI: choose the right role
It is tempting to ask which one is better and stop there. A more useful question is what role you need it to fill.
If you already have an agent stack and mainly want a voice you like, Speechify is a good place to start. If you want to evaluate an integrated real-time conversational API with tools, xAI belongs on the shortlist. If you want standalone TTS, both have an offering worth comparing.
Both are part of Nock, but that should not be read as a claim that every voice application needs both. Their public capabilities give you different ways to build. Your requirements should decide which pieces you use.
Independent benchmarks can help with the shortlist. Artificial Analysis publishes model-level voice preference comparisons, with a separate speech-to-speech evaluation for conversational systems. Those are useful evidence, not substitutes for trying your own text and tasks. They also do not prove which individual voice your users will prefer.
Does MP3 versus MP4 change the voice quality?
Usually, the file extension is not the most important decision.
MP3 is an audio codec and common file format. MP4 is a container, often holding AAC audio. MP4 is not automatically a more natural-sounding voice. The model, chosen voice, bitrate, sample rate and playback chain all matter. MDN audio codec guide
WAV and raw PCM are also worth distinguishing. A WAV file commonly wraps PCM samples with information a player can read. Raw PCM leaves your application to supply details such as sample rate and channel count. If the samples are the same, the wrapper does not make the voice more human.
For a live agent, streaming support and buffering may matter more than the file format you would choose for a download. Speechify documents MP3, AAC, Ogg/Opus and PCM streaming, with WAV reserved for its non-streaming speech endpoint. xAI's real-time API documents PCM, G.711 and Opus. Do not assume that a format accepted by a transcription endpoint is also a speech output format. Speechify formats | xAI real-time audio
My advice is to choose the voice and model first, then choose an encoding and streaming path that preserve the experience you want. A bigger file cannot fix awkward pronunciation or an unsuitable voice.
What I would test before choosing a provider
Use sentences from the product you are building, not only a polished demo script. Include a greeting, a correction, a price, an appointment time and the names your agent needs to pronounce.
Then test a few things that a short voice sample will not tell you:
- Is the voice still pleasant after several minutes, rather than just several seconds?
- Does the system pronounce names, numbers and instructions clearly?
- Can the complete agent handle interruptions and complete the right task?
- What does the whole interaction cost, including the other services it needs?
- Does the audio still sound right on the devices people will actually use?
Keep the text, loudness and playback device consistent when comparing voices. If Windows and macOS sound different, check the audio chain before assuming the provider generated a different-quality voice. Bluetooth headsets can change modes when their microphone is active, for example. Apple's Bluetooth audio explanation
That small amount of testing is more useful than treating any provider's default voice as proof of what everyone likes best.
Frequently asked questions
Are xAI and Speechify the best voice agent APIs?
They are two of our best choices for Nock because of the balance of capabilities, cost and audio quality. That is our recommendation for this kind of product, not a claim that they win every benchmark or fit every application. Speechify is primarily a TTS layer; xAI also offers a native real-time conversational API.
Do I need both to build a voice agent?
No. An integrated speech-to-speech API is one option. A modular stack with separate speech recognition, reasoning and text-to-speech services is another. Choose the components that fit your requirements rather than adding providers for the sake of it.
Is Speechify's voice quality better than xAI's?
I think Speechify's audio quality is particularly good, and it is my preference here. That is a subjective view. The fairest comparison uses the same text, comparable settings and the playback devices your audience will use. Voice quality and complete agent performance are also different questions.
Does MP4 sound better than MP3 for AI voices?
Not inherently. MP4 is a container; its audio quality depends on the codec and settings inside it. Compare the actual encoding and the source voice, not just the extension.
Why we chose them for Nock
A good voice agent should be useful enough to keep around, pleasant enough to listen to and affordable enough to use regularly.
That is why xAI and Speechify are part of Nock. They offer a practical mix of voice capabilities and published usage pricing, and I think Speechify's sound is especially strong.
If you are building a voice agent today, I would look at both. Start with the job your agent needs to do, listen to the voices on real text, and work out the full cost. Then choose the combination that makes the product better for the person using it.
Explore Nock, or read the Nock overview for how its voice, dictation and background work fit together.