tiny-audio, nanoGPT for speech-to-text

Tiny Audio A speech-to-text system you can train for $25. Tiny Audio connects a frozen, pretrained speech encoder to a pretrained LLM with a small trainable projector. The model published from this repo gets 1.8% WER on LibriSpeech test-clean and 7.4% across 12 benchmarks (11,822 samples pooled) while training only ~80M parameters. The codebase is small enough to read in an afternoon, and you can run a training loop on your laptop in about five minutes. Try it in 30 seconds No install: open the live demo, record yourself or upload a file, and get a transcript. In Python: pip install "transformers>=5.0" peft torch torchaudio librosa from transformers import pipeline pipe = pipeline( "automatic-speech-recognition", model="mazesmazes/tiny-audio", trust_remote_code=True ) print(pipe("audio.wav")["text"]) # The quarterly revenue grew by 12% according to Dr. Smith. The output is punctuated, capitalized, and has numbers formatted, with no post-processing step. The input can be a file path, a URL, or a 16 kHz numpy array. Weights are bf16, so you need roughly 6 GB of GPU or Apple Silicon memory. More than plain text # Word-level timestamps (forced alignment) pipe("audio.wav", return_timestamps=True) # {"text": "hello world", "words": [{"word": "hello", "start": 0.0, "end": 0.5}, ...]} # Who spoke when (speaker diarization) pipe("meeting.wav", return_speakers=True, num_speakers=2) Each speaker Nemotron-3-Diarization finds is transcribed separately, on a copy of the audio where everyone else is silenced, and each word belongs to the stream it came from. This is a zero-shot port of NeMo's masked_asr recipe; a word two streams both heard at once is kept once. Each speaker costs roughly their own talk time in ASR, and single-speaker audio is transcribed unmasked. Overlap is only partly handled: another person's speech inside a speaker's turn stays in that speaker's stream. Speaker diarization needs transformers installed from main (pip install git+https://github.com/huggingface/transformers) until the next release. For token-by-token streaming output, see ASRModel.generate_streaming. The model card covers batching and GPU settings. As an HTTP API on RunPod ta serve puts the model behind a batched HTTP server: requests arriving together share GPU batches, so throughput grows with load (about 460x real time at 128 concurrent requests on an RTX 4090). To run it on a RunPod GPU: poetry run ta runpod up --serve # create an inference pod; prints poetry run ta runpod wait # prints poetry run ta runpod deploy # sync the project, install the fast kernels TINY_AUDIO_API_KEY=my-secret poetry run ta runpod serve --no-attach # Ready when https:// -8000.proxy.runpod.net/health answers (a few minutes: it compiles first) Without TINY_AUDIO_API_KEY the server is open to anyone who has the URL. ta serve also runs locally, on CUDA, Apple Silicon, or CPU. Send the audio as the request body, with options in the query string: curl -X POST "https:// -8000.proxy.runpod.net/?return_timestamps=true" \ -H "Authorization: Bearer my-secret" \ -H "Content-Type: application/octet-stream" \ --data-binary @audio.wav import httpx response = httpx.post( "https:// -8000.proxy.runpod.net/", params={"return_speakers": "true", "num_speakers": "2"}, content=open("meeting.wav", "rb").read(), headers={"Authorization": "Bearer my-secret"}, timeout=600, ) print(response.json()["text"]) Options: return_timestamps, return_speakers, num_speakers and max_speakers, as in the pipeline. The response is the same dict the pipeline returns. JSON body: to send JSON instead, use {"inputs": " ", "parameters": {...}}. Audio formats: anything FFmpeg can read. Errors: 400 with {"error": ...} for bad audio or options, and 401 for a wrong key. Other endpoints: GET /health and GET /stats (batch sizes and GPU time). RunPod's HTTP proxy rejects request bodies over 500 MiB, and it drops any request that takes more than 100 seconds. For long recordings, send 16 kHz mono FLAC: ffmpeg -i recording.wav -ac 1 -ar 16000 recording.flac That's about 1 MB per minute of audio, and it costs nothing in accuracy, because the server converts everything to 16 kHz mono anyway. On an RTX 4090, 45 minutes of audio takes about 10 seconds, or 21 seconds with speaker labels. The demo Space calls the server this way. How good is it? Word error rate (%, lower is better) on 11,822 samples (up to 1,000 per dataset), measured with this repo's ta eval against the ta serve HTTP API on an RTX 4090: † Held out: no data from this source was used in training. You can check these numbers yourself and compare against commercial APIs on the same samples: poetry run ta eval -m mazesmazes/tiny-audio -d loquacious -n 100 # Same samples through a commercial API (also: deepgram, elevenlabs, apple-speech) ASSEMBLYAI_API_KEY=... poetry run ta eval -m assemblyai -d loquacious -n 100 How it works Audio (16 kHz) → speech encoder (frozen) → MLP projector (trained) → LLM decoder → Text A pretrained speech encoder turns audio into a sequence of frame embeddings. A small MLP projector stacks neighbouring frames and maps them into the LLM's embedding space. It is the only part trained from scratch. The LLM reads those projected frames as if they were tokens and writes out the transcript. Encoder, projector, and decoder are each swappable from config. Two recipes ship with the repo: Train your own Start on your laptop for free, and rent a GPU only once you know the pipeline works. git clone https://github.com/alexkroman/tiny-audio.git && cd tiny-audio poetry install # 1. Smoke test: a real training loop on your laptop poetry run python scripts/train.py +experiments=mps_smoke # 2. Before renting hardware, estimate the VRAM and disk a config needs poetry run ta runpod plan -e stage_1 # 3. Full run poetry run python scripts/train.py +experiments=stage_1 Every setting is a Hydra override, for example model.projector_hidden_dim=2048 or training.use_lora=true. When you're happy with a model, ta push publishes it to the Hugging Face Hub and ta deploy puts a demo like the one above on a Space. Training on RunPod poetry run ta runpod up -e stage_1 # create a pod with enough GPU for the config poetry run ta runpod wait # prints poetry run ta runpod deploy # sync the project and install dependencies HF_TOKEN=hf_... poetry run ta runpod train -e stage_1 poetry run ta runpod attach # watch the run in tmux Learn by building it The free 3.5-hour course walks you through the full loop: how the encoder, projector, and decoder fit together (with real tensor shapes), training a model, evaluating it against commercial APIs, and publishing it with a live demo. You need Python, the command line, and git. The course trains the smaller stage_1 recipe, not the published model, so your WER will be higher than the table above. Want to try a new projector architecture, add a dataset, or change the codebase? See CONTRIBUTING.md for the CLI reference, config layout, and quality gates. Acknowledgments Granite Speech and GLM-ASR for audio encoding Qwen3.5 and Qwen3 for language modeling LoquaciousSet for the default evaluation set License MIT