Back to the blog
Speech Applications

Ramsa Cloud Streaming: First Audio in 120 ms

An Emirati woman using a voice assistant on her phone by an office window

In cloud testing, Ramsa delivered first audio 120 milliseconds after the request was sent. That result puts a concrete number behind Ramsa Speech AI’s streaming Emirati voice: a prompt start to the speech stage of a live interaction.

For a team building an assistant, app or phone experience, the next step is to understand exactly what that interval covers. This guide explains the measurement, places published latency figures in context and gives your team a practical worksheet for evaluating real-time Arabic text to speech in its own application.

1. What the 120 ms result measures

The clock starts when the request is sent and stops when the first audio is received from Ramsa’s cloud service. It measures a request-to-first-audio interval across that cloud connection. It is a result from Ramsa’s cloud testing, rather than an estimate of how long an entire conversation will take.

Streaming delivers speech progressively, allowing playback to begin as audio becomes available. The application does not need to wait for the complete spoken reply before it can start preparing playback. That makes the beginning of the response a useful measure for interactive services.

Consider an application preparing this brief spoken opening: «جاهز؟ خلنا نبدأ.» The English meaning is “Ready? Let’s begin.” Ramsa’s role is to turn those words into Emirati speech. This is an illustrative line for an application; the 120 ms result describes the measured interval from request sent to first audio in cloud testing.

2. Give each timing boundary its own name

Several events are often grouped under “latency”, although they answer different questions. Naming them precisely lets an engineer trace the wait and lets a product manager compare the right figures.

MeasurementStart and finishWhat it tells you
Model inference timeModel processing begins to its specified audio output eventWork inside the model stage
Request to first audioClient sends the request to client receives initial audioThe boundary used for Ramsa’s 120 ms result
Request to audible playbackClient sends the request to listener hears the first soundIncludes the downstream playback path
Conversation response timeUser finishes speaking to system begins its spoken responseThe combined conversation pipeline

A complete response may also involve recognising speech, fetching information and preparing reply text before the TTS request is sent. After audio arrives, decoding, buffering and the playback destination can add time. Record those events separately when investigating a slow experience.

The phrase “time to first byte” needs the same care. A first response byte can be an envelope or metadata rather than usable audio, depending on the interface. For your test, identify the first audio payload explicitly and record audible playback as another event.

3. Build a test record your team can reuse

A useful project test begins with representative content and a written measurement definition. Include a short confirmation, a reply with names and numbers, and a longer explanation. Use inputs that resemble what your audience will actually hear rather than a single tiny phrase.

Record the following alongside each run. This is a worksheet for your evaluation, not a description of unreported conditions from Ramsa’s existing test.

  • Environment: test date, client location, service location and network path.
  • Connection: transport, connection reuse and any application proxy.
  • Speech request: exact input, voice, model identifier where available, and output format.
  • Operating conditions: first or warmed request, concurrent usage and relevant settings.
  • Events: request sent, first audio received, playback started and response completed.
  • Outcome: timing, errors, interruptions and listening observations.

Repeat the scenarios so the team can see variation. Keep unsuccessful requests in the record, and state the sample count with any summary. Report a median or percentile only when it has been calculated from the recorded observations. A complete record is more useful for planning than a timing figure detached from its conditions.

4. Read published figures on matching terms

The following vendor references were checked on 1 October 2026. They illustrate why a smaller number does not automatically describe a faster cloud experience.

ReferenceReported figureMeasurement context
Ramsa cloud testing120 msRequest sent to first audio received
ElevenLabs Flash latency documentationApproximately 75 msModel inference for typical short inputs; excludes network and application overhead
ElevenLabs Flash optimisation guide100–150 msVendor-published WebSocket TTFB examples for North America, Europe and South East Asia

The inference figure measures a different portion of the work. The regional TTFB examples are a closer reference, but still require aligned events, locations, inputs and operating conditions for a direct comparison. These are published reference points, not a matched Emirati-dialect benchmark or a market-wide speed ranking.

When building a shortlist, record the model and voice as well as the provider. Keep vendor-published values in a separate column from tests your own team runs. This gives stakeholders a clear view of what each number can support.

5. Evaluate speed and the listening experience together

A responsive Emirati voice should also carry the intended message clearly. Pair the timing review with a listening review: check dialect throughout the reply, pronunciation of names, pauses around numbers and continuity after the opening audio arrives.

Test on the actual channel. A phone integration and an app can add different playback behaviour even when they use the same speech service. If users interrupt the assistant, include that flow in acceptance testing and agree how the application should stop or replace a response.

Ramsa supports shared-cloud, organisation-cloud and on-premises deployment choices, including deployment within the UAE. Select the environment around your operating requirements, then evaluate the complete path there. The deployment conversation and the latency conversation belong in the same project brief.

Does 120 ms include generating the assistant’s answer?

No. It measures from sending the speech request to receiving the first audio. Any earlier work that prepares the reply belongs to a separate stage.

Does receiving audio mean the listener has heard it?

Receiving and playing are separate events. Measure playback on the intended device or phone channel to understand the user’s experience.

Bring a representative interaction and your intended deployment to Ramsa Speech AI. We can start with a clear result—120 ms to first audio in cloud testing—and shape an evaluation around the Emirati experience you want to deliver.

Build with Ramsa

Give your words an Emirati voice.

From live conversations to studio-quality voice-over, let your audience hear a voice that feels like home.

Talk to our team