Digisensus/parakeet-tdt-0.6b-lt-phone

Parakeet Lithuanian Phone 0.6B — 12.4 % WER on spontaneous telephone speech

A Lithuanian speech recogniser for spontaneous telephone speech, fine-tuned from nvidia/parakeet-tdt-0.6b-v3 on 426 hours of Lithuanian phone calls from the Lithuanian Phone Speech 429 h dataset. It writes readable text directly: punctuation, capitalisation, and numbers, dates, times and amounts as digits ("2026 m. rugsėjo 6 d., 9:30", "65 000 €").

Released by Digisensus, a conversational intelligence company from Vilnius, Lithuania, building custom enterprise ASR and LLM solutions for call transcription and conversation analytics.

Key finding

The LIEPA-3 spontaneous phone speech helped Parakeet a lot, but on real calls from outside that corpus the model is still far from usable.

• On held-out LIEPA-3 phone speech, fine-tuning cut the word error rate from 42.7 % to 12.4 % (normalized): about 3.4× fewer errors.
• On independent real phone calls it improved too, from 59.5 % to 39.3 %, but roughly two words in five are still wrong: not good enough for production call transcription.
• Why the gap: the LIEPA-3 phone speech consists mostly of radio and TV broadcasts in which the person speaking took part over the phone. That speech is much cleaner than regular customer-service or sales calls. The main weaknesses of LIEPA-3 as training data for real calls are:
  • generic topics: broadcast subjects, not the products, services and vocabulary of real customer conversations;
  • clean speech: fluent interview speech, not everyday callers with noise, interruptions and overlapping talk;
  • no context between segments: short, isolated clips, so the model never learns from the surrounding conversation.

Results

Word error rate, %, on clips never used in training:

Test set	parakeet-tdt-0.6b-v3 (base)	This model
LIEPA-3 phone speech, 2,000 held-out clips (2.1 h), normalized	42.7	12.4
LIEPA-3 phone speech, 2,000 held-out clips, cased + punctuated	51.5	19.5
Independent real phone calls, 488 clips (2.8 h), normalized	59.5	39.3
Independent real phone calls, 488 clips, cased + punctuated	66.8	53.4

• Normalized WER compares lowercase text with punctuation removed; cased + punctuated WER compares the text exactly as written. Both keep numbers as digits.
• Independent real phone calls: 488 human-verified clips of real Lithuanian telephone conversations from other sources. They are not public and were not used in training. They are longer and harder than LIEPA-3, so this set shows how the model does outside its training data.

Quick start

import nemo.collections.asr as nemo_asr

model = nemo_asr.models.ASRModel.from_pretrained("Digisensus/parakeet-tdt-0.6b-lt-phone")
out = model.transcribe(["call.wav"])     # 16 kHz mono audio
print(out[0].text)

Needs NVIDIA NeMo 2.x (pip install "nemo_toolkit[asr]"). Clips of up to 30 seconds match the training data; split longer recordings into segments first.

Training

• Data: 413,638 clips (425.8 h) of spontaneous telephone speech from the Lithuanian Phone Speech 429 h dataset, 0.5–30 s each, with written-form transcripts (text).
• Held-out data: before training, 3,000 clips of the same dataset were set aside and never trained on: 1,000 for validation (used to pick the checkpoint) and 2,000 for the final test. Their IDs are in eval/. These held-out sets are not the dataset's own validation/test splits, which were drawn independently and overlap with this model's training data. To reproduce the results, evaluate on eval/test_ids.txt.
• Setup: full fine-tune of the encoder and TDT decoder; 6,000 updates, effective batch 192 clips, learning rate 1e-4 with 300 warm-up steps and cosine decay to 1e-6, SpecAugment, bf16. One NVIDIA GeForce RTX 5090, 1.9 hours. Full settings are in config.json.
• Released checkpoint: step 4,500, the one with the lowest validation WER.

The numbers behind the charts are in training_log.csv and eval_log.csv.

Files

File	Content
model.nemo	Weights, configuration and tokenizer (NeMo format)
config.json	Training data, hyperparameters and validation results
training_log.csv	Loss, learning rate and gradient norm every 25 steps
eval_log.csv	Validation WER and CER every 300 steps
eval/validation_ids.txt, eval/test_ids.txt	IDs of the held-out clips in the dataset
charts/	Training loss and validation error rate

Limitations

• Domain: trained on one corpus of spontaneous phone speech. On other real calls the error rate is much higher (39.3 % normalized); expect the same for noisy audio, strong accents and specialised vocabulary.
• Dialects: regional dialect speech was not in the training data.
• Written form: the output follows the dataset's transcription conventions, including their known inconsistencies.
• Speakers: mostly adults; children and teenagers are rare in the training data.

Licence and attribution

This model is licensed under Creative Commons Attribution 4.0 International (CC BY 4.0). The full licence text is in LICENSE.

It is derived from nvidia/parakeet-tdt-0.6b-v3 by NVIDIA (CC BY 4.0) and trained on data derived from LIEPA-3 by Vytautas Magnus University, Vilnius University and the Institute of the Lithuanian Language (CC BY 4.0). When you use or redistribute it, credit all three:

Parakeet Lithuanian Phone 0.6B by Digisensus, licensed under CC BY 4.0. Fine-tuned from nvidia/parakeet-tdt-0.6b-v3 by NVIDIA (CC BY 4.0) on data derived from LIEPA-3 by Vytautas Magnus University, Vilnius University and the Institute of the Lithuanian Language (CC BY 4.0).

Citation

@misc{digisensus_parakeet_lt_phone_2026,
  title        = {Parakeet Lithuanian Phone 0.6B},
  author       = {{Digisensus}},
  year         = {2026},
  howpublished = {Hugging Face},
  url          = {https://huggingface.co/Digisensus/parakeet-tdt-0.6b-lt-phone},
  note         = {Digisensus, https://digisensus.com. Fine-tuned from nvidia/parakeet-tdt-0.6b-v3
                  (NVIDIA) on data derived from LIEPA-3 (Vytautas Magnus University, Vilnius
                  University, Institute of the Lithuanian Language), CC BY 4.0}
}