Foreword 29M

Streaming English speech-to-text for voice agents, in a 17 MB file. Foreword writes most words before the speaker has finished saying them, and it runs on one CPU thread.

  • Size: 29.2M parameters, 4-bit weights. The weights file is 16.79 MB; 17.61 MB with the config and tokenizer.
  • Base: fine-tuned from Moonshine Tiny (streaming), moonshine-ai/moonshine-streaming-tiny.
  • Speed: 12 to 13 ms of computing per 80 ms of audio on one Apple M5 thread.
  • Language: English only.

Use

pip install torch "transformers>=5" soundfile huggingface_hub
from foreword import Foreword            # foreword.py is in this folder

fw = Foreword.from_pretrained(".")       # this folder, or a Hugging Face repo id
print(fw.transcribe("speech.wav"))       # 16 kHz mono

stream = fw.stream()                     # live: feed 16 kHz mono float32 audio
for chunk in mic_chunks:                 # any chunk size
    for u in stream.feed(chunk):         # one update per 80 ms of audio
        print(u.text, u.words, u.guess)

Each update has text (the transcript so far, including the word being spoken), words (each word with a confidence), guess (up to two words the model expects next, with confidences) and time (seconds of audio heard). Earlier words can still be revised by a later update.

From the command line: python foreword.py speech.wav [--live].

Results

All figures were measured for this model with the scripts in the Foreword project.

How early words arrive. Clips are fed in 80 ms steps. A word counts as early when it is right, and never changed afterwards, before the speaker finished it. Word times come from Montreal Forced Aligner.

Test clips Words right for good before they end Median time against the word's end WER on these clips
100 LibriSpeech test-clean clips (1,870 words) 77.5% 90 ms before 4.87%
272 NOTSOFAR-1 meeting clips (4,721 words) 67.8% 80 ms before 8.17%

Accuracy on whole clips, every clip of each set (Whisper-style text normalizer):

LibriSpeech test-clean LibriSpeech test-other AMI test (re-cut) NOTSOFAR-1 eval
4.58% 13.16% 18.50% 9.86%

Open ASR Leaderboard test sets (the leaderboard's text normalizer; the whole clip in one step):

TED-LIUM VoxPopuli LibriSpeech clean Earnings22 LibriSpeech other AMI GigaSpeech SPGISpeech Mean
6.77% 12.30% 4.50% 22.42% 13.09% 18.08% 14.07% 7.79% 12.38%

The training parts of LibriSpeech, AMI and VoxPopuli are in the training data, so those test sets are familiar speech. TED-LIUM, Earnings22, GigaSpeech and SPGISpeech are not in the training data.

Against other models

Foreword is fine-tuned from Moonshine Tiny (streaming), so that is the first model to compare it with: like for like, the small file each one ships, and further down the two full-precision checkpoints in the same engine. Two other small speech-to-text models were run on the same clips too: Whistle (17 MB) and Whisper base.en (145 MB).

Streaming, each model as it ships

The same clips, fed in 80 ms steps. Each model runs the way its own project ships it:

  • Moonshine Tiny: the moonshine-voice runtime 0.1.6 with its quantized model (45 MB), re-decoding on every step.
  • Whistle: cactus-needle 3.2.10 with its engine 3.3.3 and whistle.cact (16.9 MB), English forced, in its own live mode (chunks of about 1 s); the text shown is what it has committed plus its unconfirmed tail.
  • Whisper base.en: openai-whisper 20250625 (the 145 MB checkpoint), re-run on all audio so far every 480 ms; the text shown is its latest output.

Share of words right for good before they end, on audiobook and meeting clips: Foreword 29M 77.5% and 67.8%, Moonshine Tiny 3.3% and 5.2%, Whistle 11.3% and 7.9%, Whisper base.en 29.8% and 24.8%

Median time at which a word is right for good: Foreword 29M 90 ms and 80 ms before the word ends; Moonshine Tiny 270 and 300 ms after; Whistle 390 and 490 ms after; Whisper base.en 100 and 130 ms after

Clips Model File Words right for good before they end Median time against the word's end WER on these clips
100 LibriSpeech test-clean Foreword 29M 17 MB 77.5% 90 ms before 4.87%
Moonshine Tiny 45 MB 3.3% 270 ms after 6.14%
Whistle 17 MB 11.3% 390 ms after 3.60%
Whisper base.en 145 MB 29.8% 100 ms after 3.49%
272 NOTSOFAR-1 meeting Foreword 29M 17 MB 67.8% 80 ms before 8.17%
Moonshine Tiny 45 MB 5.2% 300 ms after 11.43%
Whistle 17 MB 7.9% 490 ms after 15.88%
Whisper base.en 145 MB 24.8% 130 ms after 14.96%

Foreword minus each model on the same clips, with 95% intervals:

Words early, audiobook clips WER, audiobook clips Words early, meeting clips WER, meeting clips
Moonshine Tiny +74.3 (+71.9 to +76.6) -1.3 (-2.6 to +0.1) +62.6 (+60.5 to +64.6) -3.3 (-4.5 to -2.0)
Whistle +66.2 (+64.0 to +68.4) +1.3 (-0.1 to +2.5) +59.8 (+57.7 to +61.8) -7.7 (-9.7 to -5.8)
Whisper base.en +47.7 (+45.1 to +50.3) +1.4 (+0.2 to +2.6) +42.9 (+40.7 to +45.0) -6.8 (-11.6 to -3.5)

Foreword is the only one of the four that writes most words before they end. On the audiobook clips Whistle and Whisper base.en make fewer errors than Foreword, by about 1.3 points. On the meeting clips Foreword makes fewer errors than all three; the training part of NOTSOFAR-1 is in Foreword's training data, so that is familiar speech for it. The Open ASR sets below were not run for Whistle or Whisper base.en.

Whole clips against Moonshine Tiny: the eight Open ASR sets

Moonshine Tiny in its own runtime is given the same clip and silence as Foreword and asked for its final text.

WER differences with 95% intervals on the Open ASR sets, like for like: the 17 MB file against Moonshine Tiny as it ships, and the two full-precision checkpoints in the same engine

File TED-LIUM VoxPopuli LibriSpeech clean Earnings22 LibriSpeech other AMI GigaSpeech SPGISpeech Mean
Moonshine Tiny, its own runtime 45 MB 7.80% 15.67% 5.61% 23.64% 15.28% 26.23% 14.92% 8.20% 14.67%
Foreword 29M 17 MB 6.77% 12.30% 4.50% 22.42% 13.09% 18.08% 14.07% 7.79% 12.38%

Foreword makes fewer errors on all eight sets: 2.29 points on the mean (95% interval 2.15 to 2.43), 0.88 on the four sets that are not in its training data (0.69 to 1.06) and 3.70 on the four whose training parts are (3.50 to 3.91). The runtime returned no text for 3,364 of the 12,643 AMI clips, nearly all of them shorter than a second. Without AMI the difference is 1.45 points on the mean of the other seven sets (1.30 to 1.60).

At full precision, in the same engine: 117 MB against 176 MB

The Foreword checkpoint before its 4-bit stage (the long-full-precision variant, 117 MB) and the checkpoint moonshine-ai/moonshine-streaming-tiny (176 MB), both through the step-by-step engine Foreword runs in.

Clips Model File Words right for good before they end Median time against the word's end WER on these clips
100 LibriSpeech test-clean Moonshine Tiny checkpoint 176 MB 70.5% 70 ms before 4.39%
Foreword, full precision 117 MB 78.9% 90 ms before 3.97%
272 NOTSOFAR-1 meeting Moonshine Tiny checkpoint 176 MB 46.1% 20 ms before 13.65%
Foreword, full precision 117 MB 68.5% 80 ms before 7.93%

Foreword minus the Moonshine Tiny checkpoint: +8.4 points of words early on the audiobook clips (+6.4 to +10.4) with WER level (-0.4, -1.5 to +0.6), and +22.4 points on the meeting clips (+18.8 to +26.1) with WER 5.7 points lower (-7.8 to -3.7). Most of the earliness on read speech comes from the engine and not from the training: the Moonshine Tiny checkpoint in it is already early on 70.5% of audiobook words. What the training adds is earliness that holds on meeting speech.

Open ASR sets. SPGISpeech was not run for the full-precision Foreword checkpoint, so this is seven sets:

File TED-LIUM VoxPopuli LibriSpeech clean Earnings22 LibriSpeech other AMI GigaSpeech Mean of the seven
Moonshine Tiny checkpoint 176 MB 6.54% 12.70% 4.49% 19.71% 12.54% 18.11% 13.30% 12.49%
Foreword, full precision 117 MB 6.38% 11.40% 4.15% 20.93% 11.74% 16.57% 13.22% 12.06%

Foreword makes 0.43 points fewer errors on the mean of the seven (0.30 to 0.56). It is ahead on the four sets whose training parts are in its training data, level on TED-LIUM (-0.16, -0.59 to +0.25) and GigaSpeech (-0.07, -0.20 to +0.06), and 1.22 points behind on Earnings22 (+0.68 to +1.69).

Training

The model keeps Moonshine's streaming architecture with the word table cut to 19,007 entries. It was trained on audio cut off partway through a word, with the text so far (finishing the cut-off word) and the next two words as the target. Training ran for 40,000 steps at full precision, then 5,000 steps with the weights held to 4 bits, with meeting speech drawn twice as often in that last stage. Hesitations (uh, um) were left out of the training text.

Training data, 602,343 clips with word times:

Dataset Hours Licence
LibriSpeech 960 CC BY 4.0
People's Speech (the CC BY clean subset) 272 CC BY (its page lists versions 2.0 to 4.0)
VoxPopuli (English) 176 CC0; the recordings fall under the European Parliament's legal notice
AMI Meeting Corpus 77 CC BY 4.0
ICSI Meeting Corpus 29 CC BY 4.0
NOTSOFAR-1 (train and dev1) 16 CC BY 4.0

Licences as stated on each dataset's own page on 2026-11-19.

Limits

  • English only, 16 kHz audio.
  • Accuracy is lower on speech unlike the training data: 22% WER on earnings calls.
  • On read speech two other small models are more accurate: on the 100 audiobook clips Whisper base.en (145 MB) scores 3.49% and Whistle (17 MB) 3.60%, against 4.87% for Foreword.
  • The 17 MB file is less accurate on whole clips than Moonshine Tiny's 176 MB full-precision checkpoint: 12.38% against 11.68% on the mean of the eight Open ASR sets, 0.70 points (0.57 to 0.83). The two are level on the four sets whose training parts are in the training data (+0.03, -0.15 to +0.22), and Foreword is 1.37 points behind on the four that are not (+1.18 to +1.55). The 4-bit weights cost about one point.
  • A word shown early can change until the model settles on it. Use the confidences, or wait one update, where a wrong early word is costly.
  • Training text had no digits, and only part of it had punctuation and capitals.
  • Each model was trained once. Differences of a few tenths of a point between runs are noise.

Licence

MIT (see LICENSE). Foreword builds on Moonshine and on other MIT-licensed work; their notices are in THIRD_PARTY_LICENSES and must stay with every copy.

Downloads last month
22
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for tanzilidrisi/Foreword-29M

Finetuned
(5)
this model