OpenAI’s TTS-1-1106 text-to-speech model, accessible via OrcaRouter at zero markup with real-time streaming.
openai/tts-1-1106 is a text-to-speech model released by OpenAI in November 2023. It uses a neural network to generate speech from plain text, aiming for natural intonation, pacing, and clarity. The…
The model primarily supports English-language text, though it can process text in other languages that use the Latin alphabet with varying degrees of naturalness. OpenAI has not published an exhaustive list of supported languages, but widely known usage confirms strong performance for English, Spanish, French, German, Italian, Portuguese, and Japanese (when written in Latin script). For non-Latin scripts such as Chinese or Arabic, output quality may degrade. Users should test their target language if it falls outside the common set.
Six preset voices are available: alloy, echo, fable, onyx, nova, and shimmer. Each has a distinct tone and gender presentation (e.g., alloy is neutral, echo is masculine, fable is youthful, onyx is deep, nova is warm, shimmer is bright). The voice is selected by setting the voice parameter in the request. There is no custom voice cloning or fine‑tuning capability; you must choose from these six. For projects requiring a consistent brand voice, it may be necessary to evaluate which preset best matches the intended personality.
Yes, TTS-1-1106 supports streaming via server-sent events (SSE). By setting the stream parameter to true, the API returns audio data as a sequence of chunks, each containing a portion of the generated speech. This allows clients to begin playback almost immediately, minimizing perceived latency. Streaming is particularly useful for conversational interfaces where users expect instant verbal feedback. Note that streaming may slightly increase total bandwidth overhead, but the end‑to‑end delay is significantly lower than waiting for a full response.
TTS-1-1106 has several known limitations. First, it cannot produce singing or non‑speech vocalizations (e.g., laughter, crying) reliably. Second, the preset voices do not support emotional expression—the tone remains consistent regardless of content. Third, the maximum input text length is about 4096 characters; longer content must be split into multiple requests and concatenated. Fourth, background noise handling is absent; the model expects clean input text. Finally, performance for languages other than English may be inconsistent. For high‑quality multilingual needs, consider specialized models.
TTS-1-1106 achieves near‑human naturalness for short to medium‑length English sentences. The neural architecture produces smooth prosody, appropriate phrasing, and minimal robotic artifacts. Independent evaluations (not provided in the given facts) generally place it among the top general‑purpose TTS models available through commercial APIs. However, very long texts may occasionally exhibit unnatural pauses or mis‑stressed syllables. Overall, for most conversational and narration scenarios, the output is convincingly natural.
Latency varies with input length and network conditions, but TTS-1-1106 is optimized for low end‑to‑end delay. A typical request of 100–200 characters returns the first audio chunk in under 500 ms when streamed. Full response time for a 1000‑character input is usually under 2 seconds. These figures depend on OrcaRouter’s network and server load. Streaming reduces perceived latency significantly because playback can start before the entire audio is generated. For real‑time applications, streaming is strongly recommended.
Compared to older TTS models such as Tacotron or WaveNet, TTS-1-1106 offers faster inference and easier integration via a REST API. It generally matches or exceeds the quality of cloud TTS from Amazon Polly (neural voices) and Google Cloud Text-to-Speech (WaveNet voices) on standard English voice samples. It trails specialized models from ElevenLabs in terms of emotional range and custom voice cloning. For cost, TTS-1-1106 is priced competitively at $15 per 1M input/output tokens (with zero markup on OrcaRouter), which translates to roughly $0.015 per 1K characters of input. This is similar to Azure’s neural TTS tiers.
The main trade-offs are the limited voice selection (no customization) and the 4096‑character input cap. Users needing a brand‑specific voice must choose one of the six presets. Additionally, the model does not accept SSML tags for fine‑grained control over pronunciation or emphasis beyond basic SSML prosody elements. It also lacks support for multi‑speaker dialogue in a single request—each request produces one voice. For dialogue, you must send separate requests with different voice parameters and stitch the audio client‑side.
Pricing is based on tokens—both input and output—at a flat rate of $15.00 per 1,000,000 tokens. One token corresponds roughly to 0.75 characters of English text for input, and output tokens are counted as the equivalent number of generated audio samples (also at $15/1M tokens). There is no additional markup; OrcaRouter passes through the provider rate exactly. This pricing model means cost scales linearly with the amount of text processed and audio produced. For a typical 1000‑character input, the total cost is roughly $0.02 (input) + $0.02 (output) = $0.04 per request.
Streaming itself does not change the per‑token price; you are billed for the same number of tokens regardless of whether the response is streamed or delivered as a single blob. Caching repeated requests client‑side can save costs by avoiding duplicate API calls. OrcaRouter does not offer server‑side caching for TTS responses. If your application generates the same text often (e.g., a standard greeting), you should pre‑generate and cache the audio locally to minimize expenses. There are no volume discounts or tiered pricing; the rate is flat per token.
Consider a chatbot that reads aloud 500‑character messages. Each message would cost input: 500 chars ≈ 667 tokens → ~$0.01; output: similarly ~$0.01; total ~$0.02 per message. At 10,000 messages per month, that’s $200. For an audiobook of 50,000 words (≈350,000 characters input), cost would be roughly: input $0.07, output $0.07 → $0.14 total for the book. However, note that output token count may be higher due to audio encoding; actual costs may double. Always test a sample and use the exact token counts returned by the API to budget accurately.
No. OrcaRouter charges exactly the same price as the underlying provider (OpenAI) for TTS-1-1106: $15 per 1 million input tokens and $15 per 1 million output tokens. There are no additional fees for using the OrcaRouter gateway. This transparent pricing allows developers to access OpenAI’s TTS model without any cost increase compared to using OpenAI’s own API directly. You only pay for the token consumption, and OrcaRouter adds no extra surcharges.
You call the model using the standard OpenAI‑compatible chat completions endpoint at base URL https://api.orcarouter.ai/v1. Set the model parameter to "openai/tts-1-1106". The request body requires an input field containing the text you want to convert, and optional fields like voice, response_format, and speed (for playback rate). Authentication is done via an API key in the Authorization header. Example with curl: curl https://api.orcarouter.ai/v1/audio/speech -H "Authorization: Bearer YOUR_API_KEY" -H "Content-Type: application/json" -d '{"model":"openai/tts-1-1106","input":"Hello world","voice":"alloy","response_format":"mp3"}' > output.mp3.
Key request parameters include: model (required, set to "openai/tts-1-1106"), input (string, the text to speak), voice (string among alloy, echo, fable, onyx, nova, shimmer), response_format ("mp3" or "opus", default mp3), speed (float, default 1.0; range documented by OpenAI as 0.25 to 4.0), and stream (boolean, default false). Additional optional parameters include user, as per OpenAI conventions. Not supported: functions, tools, or logprobs. For streaming, set stream: true, and the response will be a series of SSE events with audio data.
The minimum required fields are model and input. The model must be exactly "openai/tts-1-1106". The input field must be a non‑empty string (up to ~4096 characters). If you omit voice, the API defaults to alloy. If you omit response_format, it defaults to mp3. The request must be sent as JSON over HTTPS. OrcaRouter requires a valid API key placed in the Authorization header with the Bearer schema. There is no need to specify the text_type parameter; plain text is assumed.
If you are currently using OpenAI’s own API endpoint for TTS-1-1106, migrating to OrcaRouter requires only changing the base URL from https://api.openai.com/v1 to https://api.orcarouter.ai/v1, and updating the API key to your OrcaRouter key. The model ID must be prefixed with openai/ (i.e., use "openai/tts-1-1106" instead of just "tts-1-1106"). All other parameters (input, voice, response_format, etc.) remain the same. No code changes beyond the endpoint and key are needed. Pricing remains identical, with zero markup.
OpenAI also offers TTS-1 (the original) and TTS‑1‑HD. TTS-1-1106 is an iteration optimized for lower latency and streaming, while TTS‑1‑HD provides higher quality at the cost of speed. TTS-1-1106 supports the same six voices and price point. Compared to the older TTS-1 model, TTS-1-1106 has improved naturalness and reduced artifacts, especially in longer utterances. For applications where audio quality is paramount and latency is less critical, TTS‑1‑HD may be preferable; otherwise, TTS-1-1106 is the recommended choice for interactive use.
ElevenLabs offers extensive voice cloning and emotional control, but at a higher per‑character cost (roughly $0.03 per 1K characters for their base tier). Azure Neural TTS has similar pricing to OpenAI (about $15 per 1M characters) and supports over 100 voices and multiple languages. TTS-1-1106 has a smaller voice selection but simpler API and excellent latency. If you need a custom brand voice, ElevenLabs or Azure are better; for English‑only, low‑latency, zero‑markup access via OrcaRouter, TTS-1-1106 is strong.
Cheaper TTS alternatives may exist (e.g., espeak‑ng or some open‑source models running locally on CPU) but usually lack naturalness. TTS-1-1106 is appropriate when you need cloud‑scale, high‑quality, low‑latency speech generation without managing infrastructure. For very high volume where cost is critical, consider a custom solution. If your app only needs occasional speech, TTS-1-1106’s per‑token pricing is competitive. For non‑English languages or custom voices, other models may be more cost‑effective.
TTS-1-1106’s main differentiators are its speed (sub‑second streaming) and its zero‑markup availability through OrcaRouter. It is part of the OpenAI ecosystem, so users already integrating GPT models can use the same API key and infrastructure. The six preset voices are consistent and well‑tuned for understanding. The model’s low latency makes it ideal for conversational AI, while the SSML support (basic) allows some control over pacing. However, it lacks advanced features like emotional speech, word‑level timestamps, or custom voice training – which are trade‑offs for its simplicity and speed.
OpenAI-compatible — keep the SDK you already use
https://api.orcarouter.ai/v1response_formatspeedvoice| per 1M characters | $15.00 |
|---|---|
| Currency | USD |
| This model is not billed per token. The rate above is the unit it actually meters, and it is what settlement charges. | |
What developers are saying this week
GET /api/public/models/openai/tts-1-1106Open @misc{orcarouter_tts_1_1106,
title = {openai/tts-1-1106 API},
author = {openai},
year = {n.d.},
howpublished = {OrcaRouter},
url = {https://www.orcarouter.ai/models/openai/tts-1-1106}
}openai. (n.d.). openai/tts-1-1106 API. OrcaRouter. https://www.orcarouter.ai/models/openai/tts-1-1106