n8n-nodes-nemo-speech

An n8n community node for NeMo Speech — offline speech-to-text, speaker diarization, voice-activity detection, punctuation/capitalization, inverse text normalization, and translation, powered by NeMo-Speech.cpp (NVIDIA's lightweight C++ inference runtime) and the parakeet-tdt-0.6b-v3 model (Q8_0 GGUF, ~714 MB).

Runs fully on-device / offline on CPU. No API keys, no cloud, no GPU required.

Install into n8n

n8n add-nodes n8n-nodes-nemo-speech
# or, in a container:
docker exec -it <n8n> n8n nodes:install n8n-nodes-nemo-speech

On first run the node self-heals (no postinstall):

1. downloads the static nemo-speech binary + ffmpeg to ~/.n8n/nemo-speech/bin/<platform>-<arch>/, and
2. auto-downloads the default ASR model (parakeet-tdt-0.6b-v3.q8_0.gguf) to ~/.n8n/nemo-speech/models/ on first transcription.
3. auto-downloads any enabled companion models (diarization, VAD, punctuation, ITN grammars, translation) to the same directory on first use.

No binary or model weights are shipped in the npm package. All are fetched once and cached under n8n's standard data volume (~/.n8n), so they survive container/pod restarts. To use local copies instead, set NEMO_SPEECH_PATH / FFMPEG_PATH, or point the relevant "Model Path" field at a local .gguf/directory.

How it works

1. The audio (MP3, FLAC, OGG, M4A, WAV, …) is converted to 16 kHz mono PCM16 WAV via the bundled static ffmpeg.
2. nemo-speech transcribe <wav> --device cpu runs the NeMo-Speech.cpp runtime with the Q8_0 GGUF model.
3. The transcript (plain text, JSON, SRT, or WebVTT) is returned as a JSON property.

The parakeet-tdt model runs in offline full-context mode and handles arbitrarily long audio (podcasts, lectures, meetings) — the >34-min empty-transcript bug some whisper.cpp-based tools hit is specific to that implementation, not the model.

Features

The static binary is built with NeMo-Speech.cpp's ASR core: speech recognition (ASR), speaker diarization, VAD, punctuation/capitalization (PnC), inverse text normalization (ITN), and translation (NMT). Models and companion assets are auto-downloaded from official sources on first use.

| Capability | Mode / Trigger | Notes | |---|---|---| | Speech to Text (ASR) | Operation: transcribe | parakeet-tdt-0.6b-v3 (default), nemotron-3.5-asr-streaming-0.6b, or custom HF model. | | Speaker diarization | opt-in (diarize) | Auto-downloads NVIDIA's Nemotron-3-Diarization model; tags each word with a 1-based speaker id (JSON output). | | VAD masking / endpointing | opt-in (vadMasking / vadBasedEou) | Auto-downloads the official Silero VAD 6.2.3 model. | | PnC (punctuation/casing) | opt-in (pnc) | Restores punctuation/casing for plain-text models (e.g. parakeet-ctc). | | ITN (inverse text normalization) | opt-in (itn) | "twenty twenty four" → "2024". | | Translation (NMT) | opt-in (translateTo) | Auto-downloads NVIDIA's Riva-Translate-4B-Instruct-v2 model (4.2 GB). |

Node inputs

| Input | Maps to nemo-speech | Description | |---|---|---| | Input = Binary / URL | — | Audio source: a binary property from a previous node, or a URL (wav/mp3/flac/ogg/m4a). | | Binary Property | — | Name of the binary property holding the audio (default data). | | Audio URL | positional arg | Used when Input = URL. | | Output Format | --format | text, json, srt, vtt. JSON/SRT/VTT add word timestamps. | | VAD Masking | --vad-masking | Enable VAD feature masking (auto-downloads the official Silero 6.2.3 GGUF). | | Diarize | --diarize | Enable speaker diarization (auto-downloads Nemotron-3-Diarization). | | Endpointing | --endpointing | Mid-stream end-of-utterance detection. | | Enable ITN | --itn-model-dir (when on) | Off by default. Auto-downloads the official multi-language grammars on first use. | | Enable PnC | --pnc-model (when on) | Off by default. Auto-downloads the official PnC BERT model on first use. | | Translate To | --translate-to | Target language code (e.g. es, de, fr, zh). Enables translation (auto-downloads Riva-Translate-4B-Instruct-v2). | | Write Text File | — | Also write a .txt (returned as a transcript binary property). | | Options → ASR Model | --model | Select from supported models (Parakeet TDT 0.6B v3, Nemotron 3.5 ASR Streaming 0.6B, Nemotron Speech Streaming EN 0.6B, Parakeet CTC 1.1B) or Custom. Shows Hugging Face repo ID. | | Options → Custom ASR Model | --model | Enter any Hugging Face repo ID (e.g. nvidia/parakeet-tdt-0.6b-v3) or local .gguf file path. Auto-downloads from HF. | | Options → Device | --device | cpu, cuda:0, metal, vulkan:0. Default cpu. | | Options → Stream Mode | --stream | Feed audio in 160 ms chunks through the streaming recognizer. | | Options → Word Timestamps | --word-times | Include per-word timestamps (JSON). | | Options → Verbose Logs | --verbose | Enable verbose diagnostics. | | Options → Quiet | --quiet | Suppress all logging; only the transcript is returned. | | Options → LM Path | --lm-path | KenLM .bin/.arpa for Flashlight beam search (requires a flashlight build). | | Options → Lexicon Path | --lexicon | Flashlight lexicon TSV. | | Options → Tokenizer Path | --tokenizer | SentencePiece tokenizer for OOV boosting. | | Options → Beam Size | --beam-size | Flashlight beam width (default 32). | | Options → Beam Threshold | --beam-threshold | Beam pruning threshold (default 20.0). | | Options → LM Weight | --lm-weight | LM rescoring weight (default 0.8). | | Options → Word Score | --word-score | Word insertion bonus (default 1.0). | | Options → Max Boost | --max-boost | CTC max per-word boost (default 10.0). | | Options → Chunk Size | --chunk-sec | CTC buffered window in seconds (default 0.16). | | Options → Left/Right Padding | --left-pad-sec / --right-pad-sec | CTC context in seconds (default 1.92). | | Options → VAD Model Path | --vad-model | Leave empty to auto-download the official Silero 6.2.3 GGUF when VAD masking/EOU is enabled. | | Options → VAD-Based Endpointing | --vad-based-eou | Use the VAD timeline (instead of token-silence) for endpointing. | | Options → VAD Onset/Offset/Pad | --vad-onset / --vad-offset / --vad-pad-ms | VAD thresholds and padding. | | Options → Diarization Model | --diar-model | Select Nemotron-3-Diarization (default), Diar Streaming Sortformer 4spk-v2, or Custom. Shows Hugging Face repo ID. | | Options → Custom Diar Model | --diar-model | Enter custom Hugging Face repo ID or local .gguf file path. | | Options → Profanity List Path | --profanity-list | Path to a profanity filter file. | | Options → ITN Model Dir | --itn-model-dir | Custom grammar directory; only used when Enable ITN is on. | | Options → PnC Model | --pnc-model | NVIDIA PnC BERT Base EN (default) or Custom HF repo ID / local .gguf path. | | Options → NMT Model | --nmt-model | Riva-Translate-4B-Instruct-v2 (default) or Custom HF repo ID / local .gguf path. | | Options → Speech Context | --speech-context | Comma-separated words to boost in the transcript. | | Options → Output File Base Path | — | Base path for the .txt output; only used when Write Text File is on. |

Every nemo-speech transcribe flag is exposed as a node input.

Output

Each input item produces one output item:

{
  "json": {
    "text": "And so, my fellow Americans, ask not what your country can do for you, ask what you can do for your country.",
    "model": "parakeet-tdt-0.6b-v3 (default)",
    "input": "data"
  }
}

With Diarize + Output Format = JSON, each word in text (parsed as JSON) carries a 1-based speaker field. With Write Text File enabled, a transcript binary property (the .txt) is also attached.

Companion models (auto-downloaded)

All converted from official sources and hosted at suneetk/nemo-speech-companions on Hugging Face. Downloaded once, cached under ~/.n8n/nemo-speech/models/.

| Companion | Source | Size | |---|---|---| | Diarization | nvidia/Nemotron-3-Diarization (official HF repo GGUF) | 107 MB | | VAD | snakers4/silero-vad v6.2.3 (official pip package, converted) | 1.2 MB | | PnC | NVIDIA's punctuation_en_bert.nemo (NGC, converted) | 139 MB | | ITN grammars | NeMo-Speech.cpp's official itn_configs.tar.bz2 release asset (Sparrowhawk, multi-language) | 2.3 MB | | Translation | nvidia/Riva-Translate-4B-Instruct-v2 (official HF repo, converted to Q8_0 GGUF) | 4.2 GB |

Environment variables

| Variable | Purpose | |---|---| | NEMO_SPEECH_PATH | Use this exact nemo-speech binary; skips the auto-download. | | FFMPEG_PATH | Use this exact ffmpeg binary; skips the auto-download. | | NEMO_SPEECH_MODEL_DIR | Override the model cache directory (default ~/.n8n/nemo-speech/models). |

Binary (runtime download)

The npm package contains no binaries or weights. On first run the node downloads the static nemo-speech binary (~14 MB) + static ffmpeg (~80 MB, from johnvansickle.com's static build) from a GitHub Release asset, extracted to ~/.n8n/nemo-speech/bin/<platform>-<arch>/. Both binaries are fully static (zero shared-lib dependencies), so they run on both glibc and musl (Alpine) targets.

The static binary is built from NVIDIA/NeMo-Speech.cpp source by .github/workflows/build-and-publish.yml (a new GitHub Release is cut per tag). Two small patches in patches/ adapt the upstream build for a fully static (musl-compatible) binary: static-itn-build.patch builds OpenFST/Sparrowhawk as static archives instead of shared libs, and static-itn-link.patch links the ITN target against those archives with the linker flags needed for OpenFST's static type-registration to work (--whole-archive) plus protobuf/re2/absl/zlib.

Build from source

npm install --ignore-scripts   # n8n-workflow pulls a native dep that needs no build here
npm run build                  # tsc -> dist/nodes/NemoSpeech/NemoSpeech.node.js
node smoke.test.js             # end-to-end smoke test (needs model + jfk.wav)

License

MIT. The nemo-speech binary is subject to the NeMo-Speech.cpp license (Apache-2.0). The Parakeet ASR model is subject to the CC-BY-4.0 license. Companion models retain their respective upstream licenses (NVIDIA Open Model License for Nemotron-3-Diarization, PnC, and Riva-Translate; MIT for Silero VAD).