Accuracy
93% of one-word replies heard right
Syl hears 93% of one-word yes/no replies correctly on held-out production calls.
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.
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.
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.
Step 01
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.
Step 02
Held-out calls from voice agents already in production on Bolna, not studio recordings.
Step 03
Short replies, noisy lines and mixed languages, and live on an agent it was never trained on.
Accuracy
Syl hears 93% of one-word yes/no replies correctly on held-out production calls.
Latency
The median time from the end of a reply to its transcript is 250 ms, so the agent answers without an awkward pause.
Cost
Syl is 10x cheaper than current state-of-the-art (SOTA) models.
Agent
Live on an agent it was never trained on, the agent had to ask callers to repeat 31% fewer times.
Callers
On the same agent, callers had to repeat themselves 47% fewer times, so conversations finish faster.
Phone audio
Built for narrow 8 kHz phone audio from busy streets and weak networks, and for callers who mix Hindi and English mid-sentence.
| The problem | What it means on a call | |
|---|---|---|
| Half a second of sound | A one-word reply gives the model almost nothing to guess from | The “yes” comes back empty, and the agent asks again |
| Narrow, noisy lines | Indian calls run on 8 kHz phone audio, from busy streets and weak networks | Short words get lost in the noise |
| Two languages at once | Callers mix Hindi and English mid-sentence, sometimes mid-word | A model tuned for one language misses the switch |
| They are everywhere | 6 in 10 caller replies are 3 words or fewer | 4M+ calls a month on Bolna have short replies |
| Details | |
|---|---|
| Type | Speech-to-text (ASR) model, built by Bolna |
| Built for | Short replies on phone calls: yes, no, ok, haan, nahi, ji, theek hai |
| Accuracy | 93% of one-word yes/no replies heard correctly, on held-out production calls |
| Latency | 250 ms median, from the end of a reply to its transcript |
| Cost | 10x cheaper than current SOTA models |
| In production | 31% fewer times the agent asked callers to repeat, 47% fewer times callers repeated themselves |
| Tested on | WER, SNR and turn detection evals, plus real calls on agents in production |
| Status | Live today |
| Coming next | Context aware, use-case aware and more languages |
Product
Voice agents that switch languages live when the caller does.
Product
Build and test voice agents without code, then put them on calls.
Product
Natural Hindi and Hinglish voices for the other half of the conversation.
Product
Find the calls where your agent misheard or got stuck, and fix them.
Product
The recording, transcript and extracted data for every call.
Product
Let visitors talk to your voice agent from your website or app.