Transducer

A transducer is a model that converts one sequence into another sequence. A typical example is the path from audio recording to text: thousands of measurements per second become a sentence made of letters. What’s special is the timing of the output. A transducer doesn’t wait until the recording has finished but already outputs characters while speech is happening. So it works on the fly, similar to an interpreter who stays with the speaker instead of translating only at the end of the sentence. The name comes from Latin and means roughly “one who leads across.”

Why dictation without waiting time depends on it

Anyone dictating something to a device wants to see the text appear immediately. A delay of two seconds already feels sluggish. Models that need the entire recording first cannot deliver that. The transducer was built exactly for this case and is therefore widely used in phones and assistant systems.

There is also a practical advantage: a transducer runs small enough to work directly on the device. The recording then doesn’t need to be sent to a remote computer on the internet. This saves data traffic, works in airplane mode, and is more privacy-friendly. Google's voice input on Android has used such a model directly on the phone for years.

It’s important to distinguish this from translation models like those behind modern chat programs. These first read the complete input and only then respond. For a chat, that’s not a problem, but for live subtitles it is. Transducers and such models therefore don’t compete directly but solve different tasks.

Encoder, prediction network, and the blank output

A transducer consists of three parts. The encoder processes the incoming audio and condenses short segments into an internal representation. A second network, often called the prediction network, looks only at the characters output so far and estimates what would linguistically make sense next. A third, small network combines both estimates and decides on the next character.

The actual trick lies in a special symbol, the so-called blank. It means: at this moment there’s nothing to output, keep moving through the audio. This solves a hard problem. Speech varies in speed, and nobody tells the model in advance which letter belongs to which millisecond. The blank allows audio and text to be freely shifted against each other.

During training, the model doesn’t learn a single alignment but computes over all possible alignments simultaneously. It only receives the audio recording and the finished text as input, and figures out the rest itself. This training variant with a recurrent network is called RNN-Transducer, or RNN-T for short. Today, the encoder often contains a transformer instead, but the basic idea remains the same.

From video call captions to the phone keyboard

Live captions in video conferences and on YouTube typically run on transducer models. The same applies to voice input on phones and to voice commands in cars. Call centers also use them to transcribe conversations while they are still ongoing.

In technical news, the term usually comes up in comparison with two other approaches. CTC is simpler and faster but ignores the text produced so far, making more mistakes with unusual words. Attention models are often more accurate but require the complete recording. The transducer deliberately sits in between.

A common misconception: transducers are fundamentally worse than large models like OpenAI’s Whisper. For a finished file, this is often true, since there the full context can be taken into account. In live applications, this type of model isn’t even in the running. Otherwise, one ends up comparing a real-time tool with one meant for post-processing.