Skip to content

Syl ASR model

Syl, the ASR model that hears the smallest words.

Haan, nahi, ji, theek hai. The smallest words carry the call. Syl is Bolna’s speech-to-text model, built for replies even shorter than a syllable, so your voice agents hear them on noisy phone lines and stop asking callers to repeat themselves.

A voice agent asks three questions on a phone call, the caller answers “haan”, “theek hai” and “nahi”, and Syl turns each short reply into text in 250 ms.

Syl is Bolna’s own speech recognition model, built for the shortest replies on a phone call.

  • What it is

    An ASR (speech-to-text) model from Bolna, made for utterances even shorter than a syllable: the yes, no, ok and haan ji that move a conversation forward.

  • Why it matters

    6 in 10 caller replies are 3 words or fewer, and on our calls about 1 in 10 short replies came back empty. A missed “yes” means a repeated question, or a lost call.

  • Who it is for

    Teams running voice agents on Indian phone calls, where callers answer in a word, on busy streets and weak networks, often in more than one language.

What Syl had to pass

  1. Step 01

    Evals across the board

    Word error rate (WER), accuracy in noise (SNR) and turn detection, so a win on one never cost another. A balanced model, not an average win.

  2. Step 02

    Real phone calls

    Held-out calls from voice agents already in production on Bolna, not studio recordings.

  3. Step 03

    Every slice that matters

    Short replies, noisy lines and mixed languages, and live on an agent it was never trained on.

Built for the moments that matter

  • Accuracy

    93% of one-word replies heard right

    Syl hears 93% of one-word yes/no replies correctly on held-out production calls.

  • Latency

    250 ms from reply to text

    The median time from the end of a reply to its transcript is 250 ms, so the agent answers without an awkward pause.

  • Cost

    10x cheaper

    Syl is 10x cheaper than current state-of-the-art (SOTA) models.

  • Agent

    31% fewer “Sorry, could you repeat?”

    Live on an agent it was never trained on, the agent had to ask callers to repeat 31% fewer times.

  • Callers

    47% fewer repeats

    On the same agent, callers had to repeat themselves 47% fewer times, so conversations finish faster.

  • Phone audio

    Noisy lines and mixed languages

    Built for narrow 8 kHz phone audio from busy streets and weak networks, and for callers who mix Hindi and English mid-sentence.

Why short replies are hard to hear

Most speech-to-text models are trained on full sentences. A one-word answer on a phone line gives them very little to work with.

Why short replies on phone calls are hard for speech-to-text, and what it means for a voice agent
The problemWhat it means on a call
Half a second of soundA one-word reply gives the model almost nothing to guess fromThe “yes” comes back empty, and the agent asks again
Narrow, noisy linesIndian calls run on 8 kHz phone audio, from busy streets and weak networksShort words get lost in the noise
Two languages at onceCallers mix Hindi and English mid-sentence, sometimes mid-wordA model tuned for one language misses the switch
They are everywhere6 in 10 caller replies are 3 words or fewer4M+ calls a month on Bolna have short replies

Syl at a glance

Syl ASR model: what it is, how it was tested and what comes next
Details
TypeSpeech-to-text (ASR) model, built by Bolna
Built forShort replies on phone calls: yes, no, ok, haan, nahi, ji, theek hai
Accuracy93% of one-word yes/no replies heard correctly, on held-out production calls
Latency250 ms median, from the end of a reply to its transcript
Cost10x cheaper than current SOTA models
In production31% fewer times the agent asked callers to repeat, 47% fewer times callers repeated themselves
Tested onWER, SNR and turn detection evals, plus real calls on agents in production
StatusLive today
Coming nextContext aware, use-case aware and more languages

Frequently Asked Questions

Syl is Bolna’s own speech recognition (ASR) model, built for the short replies on phone calls, such as yes, no, ok, haan, ji and theek hai. Its name comes from syllable, because it is meant for utterances even shorter than one.