dm15/parakeet-tdt-0.6b-hebrew

parakeet-tdt-0.6b — Hebrew (fine-tuned)

Hebrew ASR fine-tuned from NVIDIA parakeet-tdt-0.6b-v3 (0.6B TDT transducer). The base model covers 25 European languages but not Hebrew, so the SentencePiece tokenizer was replaced with a Hebrew BPE (1024) and the decoder/joint head reinitialized, then trained in stages. A fast, low-latency, streaming-capable Hebrew recognizer — well suited to real-time / voice-agent use.

Versions

• v4 (recommended) — parakeet-tdt-0.6b-hebrew-v4.nemo. Continued from v3 on a 2.8× larger, consensus-cleaned ~8,391 h corpus (audio-v2 filtered 7,594 h + knesset 511 h + crowd-transcribe 254 h + crowd-recital 32 h; 13.5M clips). 8×RTX 5090, staged. Improves on v3 across every benchmark — dramatically so on held-out realistic audio.
• v3 — parakeet-tdt-0.6b-hebrew-v3.nemo. ~3,000 h diverse. Podcasts test WER 13.5%.
• v2.1 / v2 / v1 — earlier ~765 h / ~254 h checkpoints (kept for reference).

Metrics — v3 vs v4 across benchmarks (greedy, identical normalization)

Benchmark	Content	v3 WER	v4 WER
ivrit-ai eval-d1	clean, directed	14.19%	7.90%
ivrit-ai eval-whatsapp	casual phone voice notes	13.82%	11.37%
FLEURS he	encyclopedic (hard vocab)	28.22%	24.71%
crowd-transcribe test (20,117)	conversational podcasts	13.25%	12.13%

WER depends heavily on the content/vocabulary of the audio, not just recording quality — the same model spans 7.9%→24.7% across sets. On clean directed speech (voice-agent-like), v4 reaches ~8% WER.

Decoding stack on the crowd-transcribe podcast test (20k clips)

Decoding	v3	v4
greedy	13.25%	12.13%
beam search (malsd_batch, beam=8)	12.72%	11.68%
beam + external-text KenLM (α=0.15)	12.00%	11.22% (CER 5.48%)

Reference: ivrit-ai/whisper-large-v3-turbo (same sets, same normalization)

Benchmark	v4 (this model)	ivrit-ai Whisper turbo
eval-d1	7.90%	4.80%
eval-whatsapp	11.37%	7.68%
FLEURS he	24.71%	18.88%

ivrit-ai's Whisper turbo is more accurate (larger 809M model, pretrained on Hebrew). This parakeet model is smaller (617M), several× faster, and supports low-latency streaming / real-time decoding, which large-v3-turbo does not — the right trade-off for voice agents. Closing the accuracy gap is feasible via knowledge distillation (re-label with Whisper turbo, retrain) and/or a larger base (parakeet-tdt-1.1b). See V4_REPORT.md for full details.

Usage (NeMo)

pip install "nemo_toolkit[asr]>=2.2" huggingface_hub
huggingface-cli login

import nemo.collections.asr as nemo_asr
from huggingface_hub import hf_hub_download
p = hf_hub_download("dm15/parakeet-tdt-0.6b-hebrew", "parakeet-tdt-0.6b-hebrew-v4.nemo")
m = nemo_asr.models.ASRModel.restore_from(p)
print(m.transcribe(["my_audio.wav"]))     # 16 kHz mono WAV -> Hebrew

Audio must be 16 kHz mono: ffmpeg -i in.mp3 -ac 1 -ar 16000 out.wav. On RTX 5090 / Blackwell, if decoding throws illegal memory access, disable the CUDA-graph decoder (use_cuda_graph_decoder=False). Best accuracy: beam search (malsd_batch, beam=8) + the external KenLM shallow fusion (α≈0.15). Note the LM (formal Wikipedia/web text) can hurt on very casual/colloquial speech — tune per target domain.

Training / provenance

• Base: nvidia/parakeet-tdt-0.6b-v3 (tokenizer + head replaced for Hebrew).
• v4 data (~8,391 h): ivrit-ai/audio-v2 (consensus-filtered: transcribe with v3, keep clips where v3 vs the Whisper pseudo-label agree, CER ≤ 0.20) + knesset-plenums + crowd-transcribe-v5
  • crowd-recital. 16 kHz mono, nikud stripped, digits→words, quality/duration filtered.
• Staged fine-tune (stage-2 48k + stage-3 15k steps), 8× RTX 5090, bf16 DDP.
• Full training source, tokenizer, manifests under training/; full report in V4_REPORT.md.

Licenses (read before production use)

• Base model: NVIDIA parakeet-tdt-0.6b-v3 — CC-BY-4.0 + NVIDIA Open Model License.
• Data: ivrit.ai License (custom) — permits commercial AI training; review the terms.

Notes on further gains

Biggest levers for lower WER: domain-matched data (a few hundred hours of the actual target audio, fine-tuned from v4) and knowledge distillation from ivrit-ai Whisper turbo. Lower-ROI: tokenizer rebuild (2048), more epochs. On general spontaneous Hebrew, ~11-12% is strong; on clean/matched audio, sub-10% is achievable (v4 is already ~8% on eval-d1).