baari

End-of-turn detection for Urdu voice agents. At every pause, baari decides whether the speaker has finished (reply now) or is only pausing mid-sentence (keep listening). baari (باری) means "turn" in Urdu.

It is Indic Smart Turn tiny, fine-tuned on Urdu phone conversations. It takes the same input and gives the same output as Smart Turn, so in Pipecat it replaces Smart Turn with one line.

📊 Results

The test set has 1,588 clips of real Urdu phone conversations from the held-out valid split of IndicVoices: 1,470 where the speaker had finished and 118 where they were pausing mid-sentence. No test speaker, and neither side of any test call, is in the training data. Because most clips are "finished", the fair score is balanced accuracy, the average of the two right-hand columns. A model that always says "finished" gets 50%.

Model	Balanced accuracy (95% CI)	Waits when unfinished	Replies when finished
baari int8 (default)	77.0% (72.9–81.0)	62.7%	91.4%
baari fp32	76.8% (72.4–80.7)	63.6%	89.9%
Indic Smart Turn tiny (int8)	73.4% (69.0–77.8)	61.0%	85.9%
Indic Smart Turn base (fp32)	70.9% (66.5–75.4)	47.5%	94.4%
Indic Smart Turn tiny (fp32)	70.0% (65.8–74.4)	51.7%	88.4%
Smart Turn v3.2 (CPU, int8)	63.4% (58.8–68.1)	57.6%	69.2%
Smart Turn v3.2 (GPU, fp32)	58.8% (54.8–63.0)	76.3%	41.3%
PuristanLabs1/urdu-turn-v2	51.8% (48.4–55.4)	16.9%	86.7%

• Scored on the CPU of a MacBook Pro (M3 Pro), with the same audio preparation Pipecat uses.
• int8 scores depend on the CPU. baari int8 gets 77.0% on Arm (Apple) and 74.9% on x86 (Intel and AMD, which most servers use). fp32 gets 76.8% on both.
• On the same clips, baari fp32 beats Indic Smart Turn tiny int8 by 3.3 points (paired bootstrap, 95% CI +0.3 to +6.4), and its own starting point, Indic Smart Turn tiny fp32, by 6.7 points (+3.8 to +10.0).

🚀 Use it

pip install baari

from baari import TurnDetector

detector = TurnDetector(threads=4)  # downloads baari-int8.onnx once
detector.is_finished(audio, sample_rate=16000)  # mono audio (float or int16) that ends at the pause
detector.probability(audio, sample_rate=16000)  # the same decision as a number from 0 to 1

In Pipecat:

from baari import model_path
from pipecat.audio.turn.smart_turn.local_smart_turn_v3 import LocalSmartTurnAnalyzerV3

turn_analyzer = LocalSmartTurnAnalyzerV3(smart_turn_model_path=model_path())

One call, including the audio preparation, on the CPU of a MacBook Pro (M3 Pro):

File	1 thread	4 threads
int8 (default)	36 ms	12 ms
fp32: TurnDetector(precision="fp32")	77 ms	23 ms

Use 1 thread on a server that handles many calls at once. baari needs 16 kHz audio, so resample 8 kHz phone audio first, for example with soxr.resample(audio, 8000, 16000).

Files, input and output

File	What it is
baari-int8.onnx	8.8 MB, the default
baari-fp32.onnx	32 MB, full precision
baari.pt	PyTorch weights, in the layout of SmartTurn in the code's train.py
metrics.json	The training record: data sizes, settings, and dev and test scores after each epoch

• Input input_features, float32 [batch, 80, 800]: the Whisper log-mel spectrogram of the last 8 seconds of 16 kHz audio, zero-padded at the front, with zero-mean, unit-variance normalisation.
• Output logits, float32 [batch, 1]: despite the name, the probability that the speaker has finished, as in Smart Turn v3. Above 0.5 means "reply now".

How it was trained

• Start: Indic Smart Turn tiny fp32, a Whisper-tiny encoder with attention pooling and a small classifier (8M parameters). It ships only as ONNX, so its weights were read back into PyTorch. A fresh model with those weights matches the ONNX file to within 6e-6.
• Urdu data: 57,182 clips from the conversations in the IndicVoices train split. Test speakers and both sides of every test call are left out.
• Mixed in: 39,925 clips from the training split of Indic Smart Turn's data, which covers 11 other Indian languages, so the model keeps what it already knew. 17,417 of the 97,107 training clips are "unfinished".
• Labels: an LLM agent read each call's transcript in order and marked whether the speaker had finished at the end of each segment, using this prompt.
• Clips: the last 8 seconds of audio ending at the segment, with the speaker's earlier segments as context. This is Indic Smart Turn's recipe.
• Training: 3 epochs, AdamW, learning rate 2e-5 with warm-up and linear decay, batch 64, fp16, on one T4 GPU in 34 minutes. Each batch weights the classes by how many of each it has, as in Smart Turn.
• Choosing the epoch: the best epoch (the 2nd) was picked on an Urdu dev set of 4,069 clips from speakers held out of training. The test set played no part in any choice.
• int8: static quantization with Smart Turn's settings: QDQ, per-channel int8 weights, uint8 activations, calibrated on 1,024 training clips.

⚠️ Limits

• The labels come from transcripts, not audio. A native speaker listened to 50 test clips. All 25 "finished" labels were right, and 19 of 25 "unfinished" labels were. So about 1 in 4 "unfinished" labels may be wrong, in the training data and in the test set, and "waits when unfinished" is a rough measure.
• Tested only on IndicVoices phone conversations. baari has not been measured on other recording conditions, or on other kinds of Urdu speech.
• In the training clips, the gaps and the last 0.2 seconds are digital silence. Real calls have background noise there.
• int8 results vary by about 2 points between CPU types. Use fp32 if you need the same answer on every machine.

Credits and license

The model is BSD-2-Clause, like the models it builds on: Indic Smart Turn (BSD-2-Clause), Smart Turn by Pipecat (BSD-2-Clause) and Whisper by OpenAI (Apache-2.0). It was trained on IndicVoices by AI4Bharat and Indic Smart Turn's data, both CC BY 4.0. The code to rebuild the data and the model is at github.com/samadpls/baari.

Made by Abdul Samad (abdulsamadsid1@gmail.com).